Monthly prediction of tropical cyclone activity over the South China Sea using the FGOALS-f2 ensemble prediction system

Shentong Li , Jinxio Li , , Jing Yng , , Qing Bo , Yimin Liu , Zili Shen ,

a State Key Laboratory of Earth Surface Processes and Resource Ecology, Faculty of Geographical Science, Beijing Normal University, Beijing, China

b State Key Laboratory of Numerical Modeling for Atmospheric Sciences and Geophysical Fluid Dynamics, Institute of Atmospheric Physics, Chinese Academy of Sciences,Beijing, China

c Southern Marine Science and Engineering Guangdong Laboratory, Guangzhou, China

d Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters, Nanjing University of Information Science and Technology, Nanjing, China

Keywords:Tropical cyclone South China Sea Monthly prediction Prediction system FGOALS-f2

ABSTRACT The monthly prediction skill for tropical cyclone (TC) activity in the South China Sea (SCS) during the typhoon season (July to November) was evaluated using the FGOALS-f2 ensemble prediction system. Specifically, the prediction skill of the system at a 10-day lead time for monthly TC activity is given based on 35-year (1981—2015)hindcasts with 24 ensemble members. The results show that FGOALS-f2 can capture the climatology of TC track densities in each month, but there is a delay in the monthly southward movement in the area of high track densities of TCs. The temporal correlation coefficient of TC frequency fluctuates across the different months, among which the highest appears in October (0.59) and the lowest in August (0.30). The rank correlation coefficients of TC track densities are relatively higher ( R > 0.6) in July, September, and November, while those in August and October are relatively lower ( R within 0.2 to 0.6). For real-time prediction of TCs in 2020 (July to November),FGOALS-f2 demonstrates a skillful probabilistic prediction of TC genesis and movement. Besides, the system successfully forecasts the correct sign of monthly anomalies of TC frequency and accumulated cyclone energy for 2020 (July to November) in the SCS.

1. Introduction

Tropical cyclones (TCs) are a type of severe meteorological disaster( Wang et al., 2013 ) that bring heavy rainfall, strong winds, and storm tides. The South China Sea (SCS), which is the largest tropical semiclosed marginal sea in the western Pacific, is affected every year by TCs during the summer and autumn seasons ( Peng et al., 2015 ). However,the characteristics of TC activity in the SCS are different from other regions of the western Pacific ( Wang et al., 2013 ; Wu et al., 2020 ). TCs generated locally or that have moved from the western Pacific both have a substantial impact on the SCS every year ( Ling et al., 2015 ; Li et al.,2019 ; Wu et al., 2020 ). During 1981—2015, roughly 9.2 TCs occurred on average in the SCS every year, of which 2.6 were locally generated and the other 6.6 were from outside (Fig. S1). Most TCs can make landfall on the coasts around the SCS, such as in southern China, the Indochina Peninsula, and the Philippines, where large human populations reside.Local TCs can generate and develop suddenly and then quickly make landfall ( Huang et al., 2016 ), while TCs moving from the western Pacific usually have stronger intensities ( Wu et al., 2020 ). Both of these two types of TCs can cause considerable damage. Therefore, it is essential to understand and accurately predict TC activity in the SCS ( Goh and Chan, 2010 ) on multiple timescales.

Fig. 1. Monthly track densities of TCs from 1981 to 2015 (July—November) in the SCS. The observations of TCs from the IBTrACS dataset in each month are shown in (a, c, e, g, i), while the ensemble-mean (from 24 ensemble members) results of TCs from the hindcast of FGOALS-f2 are in (b, d, f, g, j). The hindcast of each month was initialized on the 20th of the last month. Both the observation and the hindcast were analyzed in a 1° × 1° equidistant grid box with six-hour intervals.

Fig. 2. Rank correlation of the monthly track densities between the hindcast of FGOALS-f2 and observations of TCs in IBTrACS from 1981 to 2015 (July—November).The hindcast of each month (a—e) was initialized on the 20th of the last month. Color shading indicates the correlation coefficients that are significant at a two-sided P = 0.1 level. Grey shading indicates regions where the observed track density is non-zero for at least 25% of the years.

Fig. 3. Interannual variability of monthly TC frequency in the SCS from 1981 to 2015 (July—November). The hindcast of each month (a—e) was initialized on the 20th of the last month. The black line indicates observations of TCs in IBTrACS, while the red line indicates the average of 24 ensembles and the red area indicates the spread of one times the standard deviation of the 24 ensembles. The TCC and RMSE of the interannual variability of the monthly TC frequency between IBTrACS and FGOALS-f2 are given in the upper-right corner.

Fig. 4. Monthly observed tracks from IBTrACS (a—e), and the predicted probability of occurrence (units: %) by FGOALS-f2 (f—j) for TCs in the western Pacific from July—November 2020. The anomaly percentages (units: %) of TC frequency and ACE (k—o) in the SCS in the same period are also given. The observed tracks were classified based on the Saffir—Simpson hurricane wind scale (Tropical Storm to Category 5). The 35-member ensemble prediction of each month was initialized on the 20th of the last month. The probability of TC occurrence was analyzed in a 5° × 5° equidistant grid box with six-hour intervals.

As the transition between synoptic forecasting and seasonal prediction, monthly prediction of TCs in the SCS can better meet the current range of requirements for society. Monthly prediction of TCs focuses on the activity of TCs from many aspects, e.g., frequency, track density and intensity, at lead times ranging from between 10 and 30 days, which is considered to be a difficult timescale for forecasting ( Vitart, 2004 ).Research and operational applications in terms of monthly prediction began in the 1980s ( Miyakoda et al., 1983 ), and is still in the process of development to this day. In particular, monthly prediction has great prospects for application in the prediction of TCs in the SCS. Almost all TCs in the SCS have a lifetime shorter than 14 days ( Wu et al., 2020 ),meaning monthly prediction can cover their entire lifespan and forecast their genesis locations and tracks. Besides, the Madden—Julian Oscillation, which has a substantial impact on the genesis and movement of TCs, can be projected on the monthly timescale ( Vitart et al. 2007 ).Zhao et al. (2019) found that with reliable modulation by the Madden—Julian Oscillation for several weeks, the prediction skill for TC genesis on the monthly timescale can be significantly improved. With accurate monthly prediction of TC activity in the SCS, decision-makers in government, insurance and reinsurance companies, as well as other private sector entities, should have enough time to adopt measures to mitigate the negative impacts of TCs.

With the rapid development of dynamical subseasonal-to-seasonal prediction ( Vitart et al., 2017 ; Camargo et al., 2019 ), there have been several kinds of research carried out on the dynamical prediction of monthly TC activity. Compared to statistical prediction, dynamical prediction based on global climate models has the advantage of providing spatial and temporal patterns, which can significantly enrich the information provided by predictions of TC activity. Elsberry et al. (2010) used a dynamical monthly forecast model from the European Center for Medium-Range Weather Forecasts (ECMWF) with 51 ensemble members to make 32-day forecasts of the formation and tracks of TCs in the western Pacific and provided evidence of predictability with accurate prediction of the large-scale environment in the tropics. Gao et al. (2019) investigated the monthly prediction of hurricanes in the North Atlantic,especially major hurricanes, based on HiRAM (the Geophysical Fluid Dynamics Laboratory High -Resolution Atmospheric Model) ( Chen and Lin,2011, 2013 ; Gao et al., 2017 ; Zhao et al., 2010 ) with two grid configurations. Both configurations showed prediction skill for monthly hurricane frequency and accumulated cyclone energy (ACE), especially the 8-km nested configuration. However, the dynamical prediction of monthly TC activity and evaluation of monthly prediction skill in the SCS have not yet been studied or applied.

The motivation behind the present study is that (1) as a dynamical prediction system, FGOALS-f2 has delivered skillful prediction of TC activity on the seasonal timescale over the western Pacific and North Atlantic ( Li et al., 2019, 2021a ), but its prediction skill for marginal seas on a monthly timescale, which is an urgent societal demand, has not yet been given; (2) the real-time prediction performance for monthly TC activity in the SCS has not been verified but is essential for model development and business collaboration; and (3) considering the characteristics of TC activity in the SCS, the hazards on the coasts around the SCS from TCs, the operational applications and the requirements from society, it is meaningful and worth trying to evaluate and improve the monthly prediction skill for TCs in the SCS.

In this study, we mainly focused on TC track density. This aspect can allow us to evaluate the prediction skill from the spatial perspective and directly show the high-frequency area of TCs, both of which are significant in real-time forecasting. Besides, we were also concerned with the frequency and intensity of TCs. The interannual variability of frequency can help in evaluating the prediction skill from the temporal perspective; and the intensity, which was measured using ACE, is related to the potential severity of the damage caused by TCs. First,the monthly prediction skill for TCs in the SCS was analyzed at a 10-day lead time based on 35-year hindcasts (1981—2015) of FGOALS-f2 with 24 ensemble members. Then, the real-time monthly prediction of TCs in 2020 (July to November) was verified. The prediction system, data, and TC detection method used in this study are introduced in section 2 . The monthly TC prediction skill and real-time TC prediction results for 2020 (July to November) in the SCS are presented in section 3 , and then section 4 provides our conclusions and some further discussion.

2. Prediction system and data

2.1. FGOALS-f2 ensemble prediction system

The FGOALS-f2 ensemble prediction system was developed based on the climate system model named CAS FGOALS-f2 ( Zhou et al., 2015 ;Li et al., 2017, 2019, 2021b ; He et al. 2019 ). Since it was developed,FGOALS-f2 has performed well in the seasonal prediction of TC activity( Li et al., 2019, 2021a ), and its outputs have been submitted to CMME v1.0 (the China Multi-model Ensemble Prediction System, version 1) of the Beijing Climate Center ( Ren et al., 2019 ) and the National Marine Environment Forecasting Center of China for operational use. The horizontal resolution of FGOALS-f2 is C96 ( ～100 km), and there are four coupled components: atmosphere, ocean, land use, and sea ice.

A 35-year hindcast from 1981 to 2015 was carried out with 24 ensemble members. The atmosphere component assimilates temperature, surface wind, surface pressure, and sea-level pressure from JRA-55( Kobayashi et al., 2015 ) as the initial condition via the incremental analysis updating process ( Bloom et al., 1996 ). The ocean component uses the nudging method to assimilate the potential temperature from GODAS ( Huang et al., 2010 ) as its initial condition. The land-use and sea-ice components do not directly assimilate related data to derive the initial conditions, but they can be driven by the atmosphere and ocean components into coordination. Twenty-four ensemble members were generated based on a time-lag perturbation method ( Hoffman and Kalnay,1983 ; Kang et al., 2014 ). The real-time prediction began in early 2017,when the number of ensemble members increased to 35. Each of the members in the hindcast and real-time prediction were integrated for up to six months. The prediction frequency was once per month.

2.2. Data

We used a 35-year hindcast (1981—2015) dataset of FGOALS-f2 to evaluate the monthly prediction skill for TCs in the SCS. The initialization date was the 20th of each month, and we chose the prediction data for the next whole month ( ～10-day lead time) of the initialization to determine a monthly timescale. The interval of the output data was six hours.

The observational data for TCs were obtained from the International Best Track Archive for Climate Stewardship (IBTrACS) dataset( Knapp et al., 2010 ). IBTrACS is a multisource dataset, and the data source of observed TCs in the SCS is the China Meteorological Administration Shanghai Typhoon Institute ( Ying et al. 2014 ). The time interval of the dataset is six hours, which matches the model output.

2.3. TC detection method

Since the format of the outputs from the CAS FGOALS-f2 model is gridded data, we needed to find a way to track TCs from these outputs.Thus, we used a TC detection method similar to the method used in the GFDL climate model ( Zhao et al., 2009 ; Chen and Lin, 2013 ; Xiang et al.,2015 ), which comprised three main steps:

(1) First, we chose the local maximum 850-hPa absolute vorticity greater than 3.5 × 10−5s−1as a potential TC within a 600 km × 600 km grid box. Besides, the potential TC needed to have a minimum sea-level pressure and a warm core whose average temperature between 300 hPa and 500 hPa was 1°C warmer than that within the surrounding 2° × 2° grid box.

(2) Once the potential TC was selected, it was tracked based on step (1).If the lifetime of the whole track was longer than 72 hours and the maximum surface wind speed was faster than 15.84 m s−1, the track was judged to be a TC track.

(3) Detected TC tracks were then classified according to their surface wind speed based on the Saffir—Simpson hurricane wind scale( Simpson and Saffir, 1974 ).

3. Results

3.1. Climatology of monthly TC activity in the SCS

More than one TC appears in the SCS in each month from July to November, and the total number for the period accounts for nearly 80%of TCs in the whole year (Fig. S1). Thus, it is of great significance to analyze the climatology and prediction skill for TCs in the SCS from July to November. Fig. 1 shows the monthly track densities of TCs from 1981 to 2015 (July to November) in the SCS. Both the observation from IBTrACS ( Fig. 1 (a,c,e,g,i)) and the prediction from the ensemble-mean hindcast results of FGOALS-f2 ( Fig. 1 (b,d,f,h,j)) are shown as several 1° × 1° equidistant grids. In the observation, the area where TCs are high in density moves towards the south from July to November. However, the prediction by FGOALS-f2 shows a delay with respect to this phenomenon. The center of this high-density area in the observation moves from 20°N to 15°N during the period July—October, and reaches about 12°N in November, whereas that in the prediction by FGOALS-f2 is still near 20°N in October, before then moving to 15°N in November.( Yang et al., 2015 ) pointed out that the TC genesis locations in the SCS have a strong relationship with the Asian monsoon trough. The bias in the prediction of monthly TC activity, which may have an effect on the monthly prediction skill of FGOALS-f2, can be solved by improvement in the modulation of the large-scale patterns around the SCS.

3.2. Prediction skill for monthly TC activity in the SCS of the hindcast from 1981–2015

The prediction skill for monthly TC activity of the hindcast by FGOALS-f2 is evaluated both in spatial and temporal terms. Fig. 2 shows the spatial distribution of the rank correlation coefficients for monthly TC tracks between the FGOALS-f2 and IBTrACS data. The prediction in July ( Fig. 2 (a)), September ( Fig. 2 (c)), and November ( Fig. 2 (e)) has a more significant correlation, with the rank correlation coefficients of most of the colored areas being higher than 0.6. However, there is a lower correlation in August ( Fig. 2 (b)) and October ( Fig. 2 (d)), in which the rank correlation coefficients are generally between 0.2 and 0.6. Compared to the Brier Skill Score used in the weekly prediction of TC occurrence by the ECMWF model ( Lee et al., 2018 ), the rank correlation coefficient is a more efficient metric for evaluating the prediction skill for TCs in the SCS. Fig. 3 compares the interannual variability of the monthly frequency (July—November) of TCs from 1981 to 2015 in the SCS between the monthly prediction by FGOALS-f2 and the observational data from IBTrACS. The temporal correlation coefficient (TCC)fluctuates across the different months, with the highest TCC (reaching 0.59) appearing in October ( Fig. 3 (b)) and the lowest (at only 0.30) in August ( Fig. 3 (d)). The TCCs of other months ( Fig. 3 (a,c,e)) all range between 0.4 and 0.5. The RMSEs of these five months are similar, ranging between 1.31 and 1.72.

3.3. Real-time prediction of monthly TC activity in the SCS for 2020

In addition to the hindcast, we also carried out a monthly real-time prediction of TCs in the whole basin of the western Pacific and SCS for the year 2020. Real-time prediction, which is a significant development in numeral weather prediction of TCs in the SCS ( Peng et al., 2015 ),is key for operational applications of the ensemble prediction system.Like the hindcast, the 35-ensemble-member prediction in each month was also initialized on the 20th of the last month. Fig. 4 shows the observed tracks from IBTrACS and the results of the real-time prediction by FGOALS-f2 of TCs in the western Pacific from July to November 2020.The predicted probabilities of TC occurrence in July ( Fig. 4 (f)), August( Fig. 4 (g)), October ( Fig. 4 (i)), and November ( Fig. 4 (j)) are reasonable compared to the observation ( Fig. 4 (a,b,d,e)) in the SCS. There were no observed TCs in the SCS during the whole month of July, and the prediction shows a very low probability of occurrence in this respect.For the other three months, the areas with high probability of TC occurrence match well with the observed TC tracks in the SCS. However,the prediction in September is biased. Its area of high probability of TC occurrence lies in the northeast of the SCS, but there was only one TC,which moved westwards in the central SCS. The anomaly percentages of TC frequency and ACE in the SCS from July to November 2020 are also shown in Fig. 4 (k—o). The predicted anomaly percentages of frequency and ACE show the same sign as in IBTrACS for each month, but their amplitude is smaller.

4. Discussion and conclusion

To fulfill the requirements of society and further improve ensemble prediction systems, we evaluated the monthly dynamical prediction skill for TC activity of FGOALS-f2. More specifically, with a 35-year (1981—2015) hindcast by FGOALS-f2, we analyzed the predicted climatology of monthly TC activity and evaluated the monthly prediction skill for TC activity in the SCS. Also, we verified the performance with a real-time monthly TC prediction in the SCS for the year 2020.

The results and findings can be summarized as follows:

(1) FGOALS-f2 captures the climatology of monthly TC track densities in the SCS (July—November), though there appears to be a delay to the southward movement of the area of high track densities of TCs.

(2) FGOALS-f2 shows a certain level of prediction skill for TCs in the SCS through the hindcast. The TCCs of frequency fluctuate across the different months (July—November), with the highest (0.59) appearing in October and the lowest (0.30) in August. The rank correlation coefficients of TC track densities are relatively higher (>0.6) in July,September, and November, while those in August and October are relatively lower (0.2—0.6).

(3) The performance of FGOALS-f2 is verified via a skillful probabilistic prediction of TC genesis and movement and successful forecasting of the correct sign of monthly anomalies of TC frequency and ACE in a real-time monthly prediction of TC activity in the SCS for the year 2020.

The real-time prediction of TCs on a monthly timescale will continue to be developed and evaluated. We will submit the current version of the subseasonal-to-seasonal prediction data to the S2S prediction project( Vitart et al., 2017 ) for related research and operational applications,as well as begin to make weekly predictions ( Camp et al., 2018 ; Lee et al., 2018, 2020 ) of TCs that have recently occurred. In the future, we intend to complete our development of an initialization scheme for the ensemble prediction system, as well as enhance the horizontal resolution of the system to C384 ( ～25 km), which will help to further improve the system’s skill in predicting TC activity ( Bao and Li, 2020 ; Bao et al.,2020 ). A high-resolution downscaling method based on the dynamical prediction system will also be developed, when the horizontal resolution can be elevated to C768 ( ～12.5 km). The ultimate aim is for accurate multi-scale predictions of TCs to be developed and applied in society.

Funding

The research presented in this paper was jointly funded by the National Natural Science Foundation of China [grant number 42005117 ],the Strategic Priority Research Program of the Chinese Academy of Sciences [grant number XDB40030205 ], and the Key Special Project for the Introducing Talents Team of the Southern Marine Science and Engineering Guangdong Laboratory (Guangdong) [grant number GML2019ZD0601 ].

Supplementary materials

Supplementary material associated with this article can be found, in the online version, at doi: 10.1016/j.aosl.2021.100116 .