MJO Experimental Prediction Project:
MJO Experimental Prediction Framework
a) Model Descriptions
Near real time predictions from seven systems are being displayed. The weblinks provide information on the model and operational prediction techniques. An initial condition field helps with following the forecast evolution. Brief comments are made about each product in the context of the projects goals.- NCEP ensemble mean
- PSD Experimental ensemble-mean
- Linear Inverse Model (LIM)
- Multivariate regression model
- Lag regression model
- Coherent OLR modes
- Empirical Wave Propagation Model
1. NCEP operational ensemble-mean:
http://www.emc.ncep.noaa.gov/gmb/ens
This is based on the GFS operational model at NCEP whose characteristics are seen here (link to emc ensemble page). The initial condition is consistent with the model so spinup of the tropical precipitation field is minimal. Assimilation is 4 times daily and the model is updated/improved more or less randomly in time. The model’s climatology is unknown so systematic errors must be taken out in an ad-hoc manner (e.g., the average error over the last 30 or 60 days). The reanalysis climatology is removed to produce anomalies but this leaves systematic biases since the model has a different, unknown climatology.
2. PSD Experimental ensemble-mean:
http://www.cdc.noaa.gov/people/jeffrey.s.whitaker/refcst/week2/
This is a circa-1998 version of the old MRF that has been replaced by the GFS. The initial condition is a version of the reanalysis, which is not consistent with this version of the prediction model. As a result there is a rapid spin up of tropical precipitation in the first 12-36 hours to an approximate doubling by 36 hours. This model has been used in PSD’s reforecast project, where ensemble forecasts to 15 days were made every day for 23 years. The results have been used to provide reliable tercile forecasts of the probability distribution of temperature and precipitation over the USA at individual stations. Systematic model errors during the MJO cycle have also been determined diagnostically from the reforecast dataset.
3. Linear inverse model developed by Matt Newman of the NOAA/ESRL PSD
http://www.cdc.noaa.gov//map/lim/
This is the most sophisticated linear statistical model. Its skill is similar to full GCMs for forecasts of week 2 and beyond. This makes it useful for three purposes other than to make a forecast. Forecast skill can be predicted. Predictability studies of the MJO signal are feasible. Attribution of local anomalies can be done.
The linear inverse model (LIM) has additional “degrees of freedom” and provides forecasts of the Pacific North American pattern as well as the large-scale patterns forced by El Nino Southern Oscillation (ENSO)-induced convection anomalies. Predictability studies for 2-4 week lead times indicate that the PNA and the two circulation patterns associated with the MJO and ENSO comprise the predictable portion of the total variability. The remainder is unpredictable noise to subseasonal statistical models.
4. Multivariate regression model developed by Matt Wheeler, Australian Bureau of Meteorology:
http://www.bom.gov.au/bmrc/clfor/cfstaff/matw/maproom/RMM/index.htm
This model is well described in the link above. It has the advantage of using a seasonally varying pattern of MJO anomalies to account for the different structures between winter and summer. The pair of MJO indices provides an amplitude and phase of the MJO that can be plotted in a two-dimensional plane for monitoring purposes. Lag regression between the MJO index and other variables provides a pattern of lagged composite anomalies that can be used to make a prediction. The skill of this method is illustrated in the accompanying figures provided by Matt Wheeler. The model predicts the MJO composite anomalies as defined by the two MJO indices. It also tries to account for low frequency variations by using the last 120 days average as a persistence forecast.
http://www.icess.ucsb.edu/asr/mjo_forecasts.htm
This model projects a 20-90 day filtered data set into the future using linear regression techniques. Winter and summer have different models. The models are based on a time lagged covariance matrix of several EOFs coefficients from the filtered data. Only the MJO is predicted at least as defined by a 20-90 day filter of OLR. Further information is available in a publication and the link above.
7. Empirical Wave Propagation Model
8. PSD is in the process of accessing forecasts from additional systems.
- ECMWF
- Korea
b) Methodology
The basic approach is to collect forecasts from a number of statistical and general circulation models, plot selected forecast fields for different time averages (week 1 and week 2 for now) and evaluate the forecasts in the context of the initial state of the evolving atmospheric circulation. All the numerical models predict a quantity that can be related to the MJO, e.g., tropical precipitation, OLR or diabatic heating. These can be used to derive forecasts of the phase of the MJO, although using different variables complicates direct comparison. Most models also predict upper level subtropical streamfunction, which provides a link between tropical forcing and the extratropical circulation. Finally, forecasts of 500 mb geopotential height in the extratropics provide guidance on the position/amplitude of the ridges and troughs that affect mid-latitude weather.
Real time monitoring and a subseasonal synoptic model are also used in the forecast evaluation process. Verification statistics are computed to help gauge and monitor model forecast skill. The synoptic model is constructed from the observed MJO life cycle, the anomalous interaction of the circulation with major mountain ranges and the dispersion of wave energy in mid-latitudes. Postulated interactions or links among these different processes are part of the model.
c) The Subseasonal Synoptic Model
1) a "fast" ~1-2 day decay time scale related to the baroclinic life cycle
and synoptic-scale wave energy dispersion in mid-latitudes,
2) an "intermediate" ~6-10 day decay time scale related to teleconnection patterns, and
3) a "quasi-oscillation" with 30-60 day periods related to the MJO.
Studies of tropical-extratropical interaction have documented relationships between these mid-latitude circulation modes and the MJO, and such links are also a feature of the model described here.
The framework for the model is global atmospheric angular momentum (AAM), which varies with the amount and distribution of westerly flow and mass in the atmosphere. The MJO produces a robust signal in global AAM because as equatorial convection anomalies propagate eastward the accompanying surface wind and pressure anomalies induce anomalous mountain and frictional torques. The torques exchange angular momentum between the atmosphere and the solid earth. The MJO mountain torque comes primarily from subtropical topography and depends on the MJO's surface pressure anomalies and their placement relative to the African Highlands, the Himalayas and the South America Andes. Momentum and heat transports by atmospheric eddies organize the angular momentum produced by the torques and give rise to a zonal mean wind signal that propagates coherently poleward from equatorial regions into the subtropics. Such anomalies are strong enough to influence storm track variability and wave energy dispersion in mid-latitudes.
A sequence of the "MJO only" version of the synoptic dynamic model (SDM) is shown in the first figure above. It forms the core of the SDM. The stages are obtained through linear regression of 150 hPa streamfunction and tropical outgoing longwave radiation (OLR) on two indices of the MJO. The cyclonic streamfunction anomalies are highlighted with gray Ls, blue shading and blue lines. The anticyclonic anomalies have gray Hs and red shading/lines. The largest anomalies are lighted shaded in the appropriate color.
The OLR anomalies are a proxy for deep tropical convection anomalies. Solid black shaded areas denote increased convection and hatched areas decreased convection. The eastward movement during a cycle of the MJO can be seen in both fields. The sequence is valid during northern winter and was based on data from November-March 1979-80 to 1994-95. Note that the circulation anomalies are largest in the subtropics and relatively weak over the extratropics. This is a faithful depiction of the MJO signal when considered in isolation from other variability.
The sequence changes with the seasonal cycle. During the northern summer season, anomalies tend to shift west and north of the positions shown here and mimic the summertime basic state large scale circulation. However, a complete summer picture has not yet been constructed. For now, the seasonal base state flow is used to infer aspects of large scale circulation and OLR anomalies for the SDM during real time monitoring. In the next figure we add other transient variability to the winter sequence.
Two other types of variability have been added to the previous picture. Selected teleconnection patterns are shown schematically by the sequence of brown Ls and Hs on Stages 1 and 3. These anomalies tend to develop over several days and go through their life cycle over about 2-6 weeks. Sometimes the life cycle will coincide with an MJO since they can be forced by tropical convection. Other processes also excite them including interaction of the flow with mountains and transient eddy feedbacks. They are maintained by interaction with the mean flow. The winter SDM currently focuses on variability over the Asia- Pacific-North American region. The brown wavetrains evolve in conjunction with opposite phases of the Pacific North American and other teleconnection patterns.
Focusing on Stage 3, the linear superposition of the MJO and the teleconnection pattern favors a band of low heights in mid-latitudes stretching across the Pacific and North America. At the same time, the green Ls and Hs represent the phase of synoptic-scale disturbances that amplify, propagate east and disperse while undergoing a baroclinic-barotropic life cycle. During persistent episodes, the growth and decay tends to occur in the same preferred region or location. Typically 3-5 discrete synoptic events make up a persistent period for Stage 1 or 3, which can last on the order of weeks. The storm track tends to be more 'split' during Stage 1 and the evolution of individual synoptic events tend to be more erratic. Other features have also been added to the stages of the SDM. At Stage 4, green Hs and Ls highlight a synoptic wavetrain that connects the Indian Ocean with the US west coast. At Stage 2, the heavy gray Hs and Ls represent a medium time scale wavetrain linked to west Pacific MJO forcing. The synoptic discussion of these features is covered in a manuscript submitted to Monthly Weather Review (pdf of paper).
d) Description of Current Web Products
Two types of products are available. These include output from numerical models and periodic in depth discussions.
i) Output from numerical models (GCMs and statistical)
The following variables are being shown for the initial time and for weeks one and two:
- 500 mb height for the northern hemisphere
- Some measure of precipitation (outgoing longwave radiation
or column- integrated diabatic heating)- Upper level streamfunction
- Upper level velocity potential
These are available as total anomalies and total fields. Only the Wheeler, Jones and Wheeler/Weickmann models specifically forecast the MJO and related anomalies. Plans are to use previously described techniques to extract the MJO signal from the other models.
The following variables are shown in a Hovmoller format:
- Precipitation forecasts from the PSD ensemble mean
- Precipitation forecasts from the NCEP ensemble mean
- OLR and Coherent OLR modes, and their prediction
The Hovmoller format is being used to present preliminary, qualitative verification information. The observed OLR (leftmost figure) and the 12-hr precipitation forecasts (second figure) are used to verify the day 1-7 (third figure) and day 8-14 (fourth figure) precipitation forecasts. The contours on all figures show the coherent OLR modes as described by Wheeler and Kiladis. The latest precipitation forecast from the ensemble and coherent OLR mode forecast from the empirical model is appended out to 14 days at the end of the second Hovmoller. A 3-day running mean is applied. The other two Hovmollers show the forecasts for 1-7 days (plotted at day +4) and for 8-14 days (plotted at day +11). (Wheeler and Weickmann ) describe the forecast technique for the coherent OLR modes.
ii) The Weather-Climate Discussions
Periodically, an expert assessment of the current MJO status and short-term forecasts will be provided. These will include a discussion of possible mid-latitude climate effects from the tropics, a detailed assessment of the curent MJO status and commentary on the various intraseasonal forecasts.