Multi-model MJO forecasting during DYNAMO/CINDY period
- 3.3k Downloads
The present study assesses the forecast skill of the Madden–Julian Oscillation (MJO) observed during the period of DYNAMO (Dynamics of the MJO)/CINDY (Cooperative Indian Ocean Experiment on Intraseasonal Variability in Year 2011) field campaign in the GFS (NCEP Global Forecast System), CFSv2 (NCEP Climate Forecast System version 2) and UH (University of Hawaii) models, and revealed their strength and weakness in forecasting initiation and propagation of the MJO. Overall, the models forecast better the successive MJO which follows the preceding event than that with no preceding event (primary MJO). The common modeling problems include too slow eastward propagation, the Maritime Continent barrier and weak intensity. The forecasting skills of MJO major modes reach 13, 25 and 28 days, respectively, in the GFS atmosphere-only model, the CFSv2 and UH coupled models. An equal-weighted multi-model ensemble with the CFSv2 and UH models reaches 36 days. Air–sea coupling plays an important role for initiation and propagation of the MJO and largely accounts for the skill difference between the GFS and CFSv2. A series of forecasting experiments by forcing UH model with persistent, forecasted and observed daily SST further demonstrate that: (1) air–sea coupling extends MJO skill by about 1 week; (2) atmosphere-only forecasts driven by forecasted daily SST have a similar skill as the coupled forecasts, which suggests that if the high-resolution GFS is forced with CFSv2 forecasted daily SST, its forecast skill can be much higher than its current level as forced with persistent SST; (3) atmosphere-only forecasts driven by observed daily SST reaches beyond 40 days. It is also found that the MJO–TC (Tropical Cyclone) interactions have been much better represented in the UH and CFSv2 models than that in the GFS model. Both the CFSv2 and UH coupled models reasonably well capture the development of westerly wind bursts associated with November 2011 MJO and the cyclogenesis of TC05A in the Indian Ocean with a lead time of 2 weeks. However, the high-resolution GFS atmosphere-only model fails to reproduce the November MJO and the genesis of TC05A at 2 weeks’ lead. This result highlights the necessity to get MJO right in order to ensure skillful extended-range TC forecasting.
KeywordsMJO forecasting skill GFS, CFSv2, and UH global models DYNAMO/CINDY field campaign Air–sea coupling Atmosphere-only forecast MJO–TC interactions Extended-range TC forecasting
The Madden–Julian Oscillation (MJO) is the dominant mode of tropical convection variability on the intraseasonal timescales (Madden and Julian 1971; Zhang 2005; Lau and Waliser 2011). The MJO convective envelope initiates over the equatorial Africa and western equatorial Indian Ocean (Wang and Rui 1990a). The associated circulation systems propagate eastward as a Kelvin–Rossby wave couplet (Wang and Rui 1990b; Hendon and Salby 1994; Roundy 2012) involving multi-scale interactions (Majda and Biello 2004; Wang and Liu 2011). On its way eastward, the MJO modulates tropical cyclone (TC) activity over the Indian Ocean (Kikuchi et al. 2009; Fu and Hsu 2011), western Pacific (Liebmann et al. 1994; Nakazawa 2006), Eastern North Pacific, and Atlantic basin (Molinari et al. 1997; Maloney and Hartmann 2000; Mo 2000; Higgins and Shi 2001; Klotzbach 2010). Through upscale/downscale modulations and tropical-extratropical tele-connection, the MJO also influences global weather and climate variability (Donald et al. 2006). The recurrent nature of the MJO with a period of 30–60 days offers an opportunity to bridge the forecasting gap between medium-range weather forecast (~1 week) and seasonal prediction (longer than 1 month) (e.g., Waliser 2006; Fu et al. 2008; Brunet et al. 2010; Hoskins 2012). Most global operational and research weather/climate models, however, still face a variety of challenges to realistically simulate and accurately predict the MJO (Lin et al. 2006; Vitart et al. 2007; Wang and Seo 2009; Gottschalck et al. 2010; Rashid et al. 2010; Fu et al. 2011; Weaver et al. 2011; and Matsueda and Endo 2011), therefore, severely hindering the extended-range TC forecasting (Belanger et al. 2012; Fu 2012) and the prediction of MJO’s global impacts (Vitart and Molteni 2010).
Since the late August 2011, NCEP CPC has been preparing a MJO-discussion-summary3 for DYNAMO/CINDY field campaign each week that assembled two-week-lead forecasts from at least seven operational centers. After carefully reviewing all forecasts from September 1st, 2011 to March 31st, 2012, we have highlighted the major strength and weakness of these operational forecasts in a supplementary material.4 Current operational intraseasonal forecasting systems have higher skill for the successive MJOs-I, II, and V than that for the primary MJO in September and the MJOs-III and IV. During the entire DYNAMO/CINDY period, the common problems of operational models on MJO forecasts are: (1) the failure to predict the September primary MJO even with 1 week lead; (2) too slow eastward propagation; (3) the Maritime Continent barrier; (4) the difficulty to predict the MJO initiated by Rossby (or mixed-Rossby-gravity) waves; and (5) the underestimation of the observed intensity.
This study aims to: (1) quantify the MJO forecasting skills of NCEP operational models and UH (University of Hawaii) research model during DYNAMO/CINDY period; (2) reveal the strength and weakness of the models on forecasting initiation and propagation of the MJO; (3) assess the impacts of air–sea coupling on MJO forecasting and possible consequence on extended-range TC forecasting. The outline of this paper is as follows. In Sect. 2, we briefly introduce the three state-of-the-art global models: the NCEP GFS (Global Forecast System) atmosphere-only model, CFSv2 (Climate Forecast System version 2) and UH coupled models, along with the data and methodology used in this study. Section 3 documents the overall MJO forecasting skills during the entire DYNAMO/CINDY period and that just during the IOP in three models and highlights the strength and weakness of individual models in forecasting the initiation and propagation of the MJO. In Sect. 4, the impacts of air–sea coupling, forced with persistent, forecasted and observed daily SST on MJO forecasting skills are assessed through a series of forecasting experiments using UH model. Section 5 examines three models’ capability in representing the interactions of November-MJO and Thanksgiving-TC and potential impacts on extended-range TC forecasting in these models. Discussions and concluding remarks are given in Sect. 6.
2 Models and methodology
This study analyzes forecasts from three models: The Global Forecast System (GFS), the Climate Forecast System version 2 (CFSv2), and the University of Hawaii (UH) model. The GFS used to generate forecasts for the DYNAMO/CINDY period is the NCEP operational 2-week atmosphere-only forecasting system. The horizontal resolution is T574 (~27 km) for the first week and reduces to T190 (~70 km) for the second week. This system is initialized with the NCEP Global Data Assimilation System (GDAS). The forecasts are driven by the observed climatological SST plus initial SST anomaly that decays with lead time at an e-folding time scale of 90 days. This setting results in a basically persistent SST forcing during the 2-week forecast period. Daily four-time (00Z, 06Z, 12Z and 18Z) forecasts are treated as 4 ensemble members and are averaged together to get daily ensemble-mean forecasts. As an operational system, the GFS undergoes continued changes in both model physics and initialization system (GDAS), which can be tracked online (at http://www.nco.ncep.noaa.gov/pmb/changes/). The CFSv2 is the latest generation of the Climate Forecast System at NCEP, which became operational in March 2011 (Saha et al. 2013). The atmospheric component is the GFS version as of May 2007 with a horizontal resolution of T126 (about 100 km). The ocean component is the MOM4. The CFSv2 includes a comprehensive land model and sea ice model. It is initialized from the NCEP Climate Forecast System Reanalysis (CFSR, Saha et al. 2010). The 45-day forecasts are available four times (00Z, 06Z, 12Z, and 18Z) a day with four ensemble members at each time, forming an ensemble of 16 members for each day. The daily average of the 16 members has been used as ensemble mean in following analysis.
The UH model is an atmosphere–ocean coupled model (Fu et al. 2003) developed at International Pacific Research Center, University of Hawaii at Manoa. The atmospheric component is a general circulation model (ECHAM-4) with T106 resolution (about 125 km) that was originally developed at the Max Planck Institute for Meteorology, Germany (Roeckner et al. 1996). The Tiedtke-Nordeng mass flux scheme (Tiedtke 1989; Nordeng 1994) is used to represent the deep, shallow, and midlevel convections. The ocean component is an intermediate upper-ocean model developed at University of Hawaii. It is comprised of a mixed-layer and a thermocline layer with a horizontal resolution of 0.5 × 0.5-degree. The UH model carried out 45-day forecasts during DYNAMO/CINDY period each week initialized with final operational global analysis on 1 × 1-degree produced by NCEP, also known as FNL (at http://dss.ucar.edu/datasets/ds083.2), which is almost the same as GDAS analysis but generated 1 h later.
In order to assess the MJO forecasting skills in these three models during the DYNAMO period, the anomalies of observed and forecasted OLR, zonal winds at 850 and 200-hPa are first obtained by removing observed climatological annual cycle (mean plus first three harmonics). Interannual anomalies represented with most recent 120-day mean (Lin et al. 2008; Gottschalck et al. 2010) are also removed. NOAA satellite OLR and FNL winds have been taken as the observations in this study. Combined anomalies are projected onto Wheeler–Hendon’s EOF1 and EOF2 to get the time series of RMM1 and RMM2 (Wheeler and Hendon 2004); then we calculate the bivariate anomaly correlation coefficient (ACC) and root-mean-square error (RMSE) of the forecasts as a function of forecast lead time during this period (Lin et al. 2008; Gottschalck et al. 2010).
3 MJO forecasting skills in three models
Because the atmospheric component of the CFSv2 is similar to the operational GFS, at the same time, the GFS and CFSv2 forecasts used initial conditions from similar initialization systems (GDAS and CFSR), the skill differences between them may be largely attributed to the impact of air–sea coupling. Initially, these two forecasts have almost the same skill (with a high correlation of 0.96) with the skill of the GFS slightly higher in first week. The ACC of the GFS forecasts, however, falls much more rapidly than that of the CFSv2 with increased forecast lead time (Fig. 2). The present result suggests that air–sea coupling extends MJO forecasting skill by about 1 week. This finding is basically consistent with the result from our previous predictability study of Fu et al. (2007), which showed that air–sea coupling extends monsoon intraseasonal predictability by 1 week.
On the other hand, since both the CFSv2 and UH models include two-way air–sea interactions, the differences between them are largely attributed to different model physics. In first week, both the CFSv2 and GFS have higher skills than the UH model during the IOP (Fig. 3), but the UH model is consistently better beyond 10 days. This behavior of the UH model suggests that the initial conditions generated by a foreign model do cause some initial shocks, resulting in lower skills than those models (here, GFS and CFSv2) with initial conditions generated by more consistent initialization systems (GDAS and CFSR). The better skill of the UH model after first 10 days during the IOP, therefore, should be attributed to more realistic representation of intrinsic MJO mode in the model. This result is consistent with previous findings that ECHAM-4 family coupled models have an intrinsic MJO well mimic the observed one (Kemball-Cook et al. 2002; Fu and Wang 2004; and Kim et al. 2009).
4 Important role of air–sea coupling
Forecasting experiments with UH model under different SST settings
Names of experiments
Atmosphere–ocean coupled forecasts
Fcst_SST (or fsst)
Atmosphere-only forecasts driven by daily SST derived from the ‘cpl’ forecasts
Pers_SST (or psst)
Atmosphere-only forecasts driven by persistent SST
TMI_SST (or osst)
Atmosphere-only forecasts driven by observed daily TMI SST
5 November-MJO and Thanksgiving-TC
The possible causes for this bias in UH model are two folds. First, a detailed examination of each ensemble indicates that the model tends to produce too much tropical cyclone-like disturbances in southwest Indian Ocean. This bias is also present in boreal summer, which leads to an early false onset of northward-propagating boreal-summer monsoon intraseasonal events (Fu et al. 2013). Second, the bias may be related to the not-so-strong divergence and dry air intrusion associated with Rossby-wave-like response to the MJO convection, which is supposed to induce large boundary-layer divergence and bring subtropical dry air into the backside of the MJO convection and to shut off deep convection there, thus helping move the MJO eastward (Matthews 2000). Further research to unravel the detailed causes of this model bias is warranted, which will be one of our future research topics.
6 Discussions and concluding remarks
The observed SST-precipitation quadrature phase relationship on intraseasonal timescales (Shinoda et al. 1998; Senguta et al. 2001) has suggested that interactive air–sea coupling represents an important process for MJO dynamics (Wang and Xie 1998). On the one hand, the atmospheric forcing of the MJO changes underlying SST through modifying surface heat fluxes (Senguta et al. 2001; Waliser et al. 2003), oceanic mixed-layer entrainment (Fu et al. 2003; Saji et al. 2006), and horizontal advection (Han et al. 2007). On the other hand, the resultant intraseasonal SST anomaly feeds back to organize MJO convection and associated circulations through enhancing boundary-layer convergence (Wang and Xie 1998; Waliser et al. 1999) and surface evaporation (Fu et al. 2008). Interactive air–sea coupling is also found to be a necessity to maintain the observed SST-precipitation quadrature phase relationship while the forced atmosphere-only simulations produce an in-phase intraseasonal SST-precipitation relationship (Wu et al. 2002; Fu et al. 2003; Zheng et al. 2004; Matthews 2004). This result might be interpreted as that atmosphere-only approach is not an appropriate way to carry out MJO forecasting.
Our forecasting experiments, however, demonstrate that this is not the case for the well initialized runs. In fact, the atmosphere-only forecasts driven by daily SST derived from the coupled forecasts reach a similar skill level as the coupled forecasts (Fig. 6). Like the coupled forecasts, the initialized atmosphere-only runs are also able to maintain the observed SST-precipitation quadrature phase relationship to some extent (e.g., Fig. 8). It is the match between MJO-related large-scale circulations in the initial conditions and specified underlying SST sustains the observed quadrature SST-precipitation relationship in the atmosphere-only forecasts (e.g., Figs. 15 and 16 in Fu et al. 2008). The atmosphere-only free simulations forced with daily SST, however, are not an appropriate way to assess the impacts of intraseasonal SST forcing due to the mix-up of atmosphere internal MJO mode and SST-forced intraseasonal response.
The above findings raise the possibility to first improve the forecasts of intraseasonal SST anomalies through improving individual coupled models or developing multi-model ensemble; then using the resultant intraseasonal SST anomalies as boundary conditions to force atmosphere-only forecasts. Along this line, we expect that the high-resolution GFS driven by daily SST forecasted from the lower-resolution CFSv2 can reach a much higher MJO forecasting skill than its current level as forced by persistent SST. This could be true for all two-week extended-range forecasting systems participating in the THORPEX TIGGE program. Further if the improved MJO forecasting in these systems also results in better extended-range TC forecast, using forecasted daily SST as sea surface conditions will be an approach worth being implemented for all TIGGE forecasting models.
6.2 Concluding remarks
During the DYNAMO/CINDY field campaign, five MJO events have been observed (Fig. 1). The associated convective envelope resides over Indian Ocean, respectively, in the late October (MJO-I), late November (MJO-II), late December (MJO-III), late January (MJO-IV), and late February-early March (MJO-V). The convective envelopes of the two events in December and January have limited longitudinal extent. Whether they should be categorized as MJO is still an open issue that, however, won’t affect our conclusions. Current models have relatively higher skill in forecasting the MJO that follows a preceding event (successive MJO) than that without preceding event (primary MJO). The common model problems include too slow eastward propagation, the Maritime Continent barrier, the difficulty to predict the MJO initiated by Rossby (or mixed-Rossby-gravity) waves, and the underestimation of the observed intensity.
The MJO forecasting skills of the GFS, CFSv2 and UH models during the DYNAMO/CINDY period have been assessed with the Wheeler–Hendon bivariate ACCs and RMSEs (Lin et al. 2008). The overall MJO skills for the three models are, respectively, 13, 25, and 28 days (Fig. 2) when the ACC dropping to 0.5 has been used as the criterion. Two case studies have been given to show that the initiation of November MJO can be predicted with a lead time of 2 weeks by both the CFSv2 and UH models (Fig. 4). The relatively lower skill of the CFSv2 than the UH model may be largely attributed to the slow MJO eastward propagation in the CFSv2 forecasts (Fig. 5).
The superior performance of the CFSv2 over the GFS (Fig. 2) suggests that air–sea coupling significantly extends MJO forecasting skill. In order to quantify the impacts of air–sea coupling, three more sensitive forecasting experiments in addition to the coupled forecasts (Table 1) have been carried out with the UH model: atmosphere-only runs forced by persistent SST, forecasted and observed daily SST. Because initial conditions of these runs are the same, the skill differences among them directly measure the impacts of air–sea coupling and different SST settings. It turns out that the case driven by persistent SST has the lowest skill of 20 days (Fig. 6); the case driven by forecasted daily SST has a similar skill as the coupled forecasts (28 days); the case driven by observed daily TMI SST reaches beyond 40 days.
The UH and CFSv2 coupled models are also superior over the high-resolution GFS atmosphere-only model in representing the interactions of November MJO-Thanksgiving TC on extended-range timescales. The CFSv2 and UH coupled models are able to capture the development of westerly wind bursts associated with November-MJO and the genesis of Thanksgiving-TC with 2 weeks’ lead while the GFS totally fails (Figs. 10, 11), although apparent biases still exist in the forecasted MJO by the CFSv2 and UH models (e.g., intensity and spatial pattern) and the location of the TC (Fig. 11). Both the CFSv2 and UH coupled models are also able to maintain November MJO as in the observations after Thanksgiving-TC moving away, while the MJO in the GFS disappears (Figs. 12, 13).
In order to further advance our understanding of the physical processes governing the initiation and propagation of the MJO and to improve their representations in state-of-the-art global models, more in-depth diagnostics and numerical experiments are needed to address following questions: (1) How does air–sea coupling extend the MJO forecasting skills? (2) What are the major air–sea coupling processes misrepresented in the CFSv2 and UH coupled models? (3) Why does the MJO in the CFSv2 propagate so slow? (4) Why is the UH model prone to false TC genesis in southwest Indian Ocean? (5) What are the different physical processes leading to the initiations of the primary and successive MJO events? Some of these questions are under investigation and findings will be reported elsewhere.
Fishermen died unnecessarily in Tropical Cyclone-TC05A: http://maddenjulianconversation.blogspot.com/2011/11/fishermen-died-unnecessarily-in.html.
All real-time forecasts and discussions are available online: http://catalog1.eol.ucar.edu/cgi-bin/dynamo/report/index).
This material is available online: http://www.soest.hawaii.edu/~xfu/dynamo_op_fcst.pdf.
JTWC stands for Joint Typhoon Warning Center; more details of “TC05A” can be found online: http://en.wikipedia.org/wiki/2011_North_Indian_Ocean_cyclone_season#cite_note-47.
For example, the study of Jim Moum at: http://www.eol.ucar.edu/projects/dynamo/meetings/2012/jul/index.html ).
This work was sponsored by NOAA (NA11OAR4310096 & NA10OAR4310247), NSF (AGS-1005599) and by the Japan Agency for Marine-Earth Science and Technology (JAMSTEC), NASA, and NOAA through their supports of the IPRC. Additional supports are from APEC Climate Center and CMA project (GYHY201206016). We thank Dr. Matt Wheeler for sharing his codes to filter out MJO and equatorial waves in Fig. 1. This paper is SOEST contribution number 8961 and IPRC contribution number 992.
- Fu X (2012) Extended-range TC forecasting: opportunity and challenge. US CLIVAR summit, Newport Beach, CA, July 17–20, 2012. Available online at: http://www.usclivar.org/sites/default/files/meetings/Fu_ExRangeTCF_2012Summit_pres.pdf
- Fu X, Wang WQ, Lee J-Y, Wang B, Vitart F (2013) Intraseasonal forecasting of Asian summer monsoon in four operational and research models. J Clim 26:4186–4203Google Scholar
- Gottschalck J, Wheeler M, Weickmann K, Vitart F, Savage N, Lin H, Hendon H, Waliser D, Sperber K, Nakagawa M, Prestrelo C, Flatau M, Higgins W (2010) A framework for assessing operational Madden-Julian oscillation forecasts: a CLIVAR MJO working group project. Bull Am Meteorol Soc 91:1247–1258CrossRefGoogle Scholar
- Kubota H, Yoneyama K, Jun-Ichi H (2012) Contribution of tropical cyclone for the preconditioning of the Madden-Julian oscillation during CINDY2011. Presentation on AGU fall meeting, San Francisco, December 03, 2012Google Scholar
- Lau WKM, Waliser DE (eds) (2011) Intraseasonal variability of the atmosphere–ocean climate system, 2nd edn. Springer, Heidelberg, p 613Google Scholar
- Liebmann B, Hendon HH, Glick JD (1994) The relationship between tropical cyclones of the western Pacific and Indian Oceans and the Madden-Julian oscillation. J Meteorol Soc Japan 72:401–412Google Scholar
- Nordeng TE (1994) Extended versions of the convective parameterization scheme at ECMWF and their impact on the mean and transient activity of the model in the tropics. Technical Memorandum No. 206, European Centre for Medium-Range Weather Forecasts, Reading, United KingdomGoogle Scholar
- Roeckner E et al. (1996) The atmospheric general circulation model ECHAM-4: model description and simulation of present-day climate. Max-Planck-Institute for Meteorology Rep. 218, p 90Google Scholar
- Saha S et al (2013) The NCEP climate forecast system version 2. J Clim submittedGoogle Scholar
- Straub KH (2013) MJO initiation in the real-time multi-variate MJO index. J Clim 26:1130–1151Google Scholar
- Waliser DE (2006) Predictability of tropical intraseasonal variability. In: Palmer T, Hagedorn R (eds) The predictability of weather and climate, 1st edn. Cambridge University Press, Cambridge, pp 275–305Google Scholar
- Wang WQ, Seo K-H (2009) The Madden-Julian oscillation in NCEP coupled model simulation. Terr Atmos Ocean Sci 20(5):713–725. doi: 10.3319/TAO.2008.09.17.01(A)
- Wang WQ, Hung M-P, Weaver SJ, Kumar A, Fu XH (2013) MJO prediction in the NCEP climate forecast system version 2. Clim Dyn (in press)Google Scholar
Open AccessThis article is distributed under the terms of the Creative Commons Attribution License which permits any use, distribution, and reproduction in any medium, provided the original author(s) and the source are credited.