Extended Range Prediction of Monsoon Intraseasonal Variability (ISV)

1
Extended Range Prediction of Monsoon
Intraseasonal Variability (ISV)

• Matthew Wheeler
• Climate Forecasting Group
• Bureau of Meteorology Research Centre
• Melbourne, Australia.

2
Acknowledgements
Harry Hendon, Duane Waliser, Klaus Weickmann,
Xianan Jiang, Harun Rashid, and Nicholas Savage
contributed to this presentation. And I thank
the local organising committee for this
invitation.
3
Introduction
Recent research has started to fill the gap that
has traditionally existed between weather (days)
and climate (seasons) prediction. E.g. Waliser
et al. (1999), Lo and Hendon (2000), Wheeler and
Weickmann (2001), Goswami and Xavier (2003),
Webster and Hoyos (2004), and review by Waliser
(2005). Due to the difficulties in accurately
representing ISV in dynamical models, however,
most of this research and development has been
with empirical schemes. Also, due to the
prominence of the MJO on this time scale, it is
the MJO that has received most attention,
especially, but not exclusively, in austral
summer.
4

• Here, I thus also concentrate on the MJO. In
  particular, I will
• Discuss two approaches to empirical prediction
  at BMRC,
• Suggest a statistical benchmark for MJO
  prediction, and
• Examine a MJO diagnostic for dynamical forecast
  models.
• The work presented draws upon, and provides input
  to, two international working groups
• The Experimental MJO Prediction Project
  http//www.cdc.noaa.gov/MJO/
• U.S. CLIVAR MJO Working Group http//www.usclivar
  .org/Organization/MJO_WG.html

5
Two approaches to empirical prediction at BMRC

• a) Wavenumber-frequency filtering (very briefly)
  
• http//www.bom.gov.au/bmrc/clfor/cfstaff/matw/map
  room/OLR_modes/
• (operating for 7 years)
• b) Projection of daily observations onto combined
  EOFs of OLR, u850, and u200 to get two indices -
  what we call Real-time Multivariate MJO (RMM) 1
  and RMM2.
• http//www.bom.gov.au/bmrc/clfor/cfstaff/matw/map
  room/RMM/
• (operating for 4 years)

6
a) Wavenumber-frequency filtering of OLR
1/ Monitoring and forecast from 16th January
2/ Monitoring and forecast from 5th February (20
days later)
16th Jan
As described by Wheeler and Weickmann (MWR, 2001)
7
Are these MJO prediction relevant to the monsoons?
Correlation skill in Southern Summer for Day 15
Correlation skill in Northern Summer for Day 15
8
b) The Real-time Multivariate MJO (RMM) Index
Index described by Wheeler and Hendon
(2004). Statistical forecasts with index
described by Maharaj and Wheeler (2005) and Jiang
et al. (2007). The idea is that by projecting
daily observed data (with long-time scale
components carefully removed) onto the MJOs
multivariate spatial structure, you can isolate
the signal of the MJO without the need for a
band-pass time filter. EOFs of the combined
fields of 15S to 15N-averaged OLR, u850, and
u200, for all seasons.
9
Madden and Julian (1972)
The pair of EOFs describe the convectively-coupled
vertically-oriented circulation cells of the
canonical MJO, as is detectable in all seasons.
10
It is thus convenient to view the state of the
MJO in the two-dimensional phase space defined by
the two EOFs. For example, looking at the 40
days up to 17th July 2007. We can use this index
for empirical prediction.
11
Can form season-specific MJO composites using the
seasonally-independent index.
JJA
RMM phase space for all days in JJA from 1974 to
2006.
Approximately 200 days in each phase
DJF
RMM phase space for all days in DJF from 1974 to
2006.
12
MJO composite for JJA
Reproduces some (1/2) of the northward
propagation in the Indian monsoon.
13
MJO composite for DJF
Reproduces the southward excursion of convection
into northern Australia.
14
Example MJO impact on rainfall (DJF) Probability
that the weekly rainfall accumulation will exceed
the upper tercile.
15
Can forecast RMM1 and RMM2 values using multiple
linear regression
RMM1(lag)a1b1?RMM1(0) c1?RMM2(0) RMM2(lag)a
2b2?RMM1(0) c2?RMM2(0) where a1,a2,b1,b2,c1,
and c2 are computed independently for each lag,
and are a smoothly varying function of the time
of year.
Example from 17th July
15 day forecast (for 1st August)
Skill as measured by the correlation coefficient
? 0.5 for a 15-day forecast (Maharaj and Wheeler
2005)
16
Similarly, can forecast any field using
seasonally-varying, lagged, multiple linear
regression against RMM1,RMM2 at Day 0.
OLR and 850hPa wind anoms
Initial Condition on 17 July, 2007
15-day forecast for 1st August
17
Skill of RMM-based predictions Correlations of
the predicted OLR anomalies with observed OLR
(100-day high-pass)
15-day forecasts Southern Summer
15-day forecasts Northern Summer
Only modest skill (like other empirical schemes),
but a useful benchmark for dynamical MJO
forecasts (Jiang et al. 2007)
18
MJO forecast skill as a function of MJO phase Is
there a statistical MJO predictability barrier?
Less skill for forecasts going through Phase
2. But drop in skill is only relatively minor.
Figure courtesy of Xianan Jiang
19
A MJO diagnostic for dynamical forecast models A
current activity of the US-CLIVAR MJO Working
Group (Waliser/Sperber) (http//www.usclivar.org/O
rganization/MJO_WG.html) Project numerical model
forecast data onto the same two RMM EOFs that
were derived from observations.
EOFs of the combined fields of 15S to
15N-averaged OLR, u850, and u200
20
Projection of daily analyses/model forecasts onto
EOFs provides a diagnostic with which to measure
the state of the MJO in both observations and
model forecasts.
Two examples Obs empirical forecast
Obs dynamical forecast 1
1st August
1st August
Met Office Global and Regional Ensemble
Prediction System
Daily-updated dynamical forecasts available from
http//www.cdc.noaa.gov/MJO/Forecasts/index_phase.
html
21
Projection of daily analyses/model forecasts onto
EOFs provides a diagnostic with which to measure
the state of the MJO in both observations and
model forecasts.
Two examples Obs empirical forecast
Obs dynamical forecast 2
1st August
1st August
1st August
NCEP Global Ensemble Prediction System
22
Latest ten forecasts from the Bureau of
Meteorologys coupled model (POAMA). Initial
conditions for each forecast are 1 day apart (as
labelled, starting on 9th July). Coloured lines
show the trajectories of 30-day forecasts, with a
black dot placed every 5 days. Ensemble mean
(blue curve) is mean of just the last 5
forecasts.
15th August
1st August
23
What have the observations done up until
yesterday? In this case, the dynamical models
appeared to have performed reasonably well, as
the observations have reproduced the relatively
slow eastward propagation.
24
But how skilful are these dynamical forecasts in
a general sense? Our benchmark statistical
forecast is provided by lagged linear regression
with RMM1 and RMM2 as predictors. Correlation
skill of this benchmark is 0.5 for a 15-day
forecast (Maharaj and Wheeler 2005) For POAMA we
have a comprehensive hindcast dataset with which
to assess its skill (10-member ensemble started
on the 1st of each month during 1980-2005).
25
POAMA example hindcasts and observations.
When assessed using all seasons, correlation
skill 0.45 at 15 days! A little less than our
statistical benchmark, despite POAMA having a
better than average MJO.
26
For predictions of the total anomaly field (i.e.,
not just the MJO component), dynamical models
look a little better.
Experimental MJO Prediction Program
(http//www.cdc.noaa.gov/MJO/)
NCEP ENS Ensemble mean of 2004 version of the
GFS operational model (T254 L64). CDC ENS
Ensemble mean of forecasts from frozen version of
MRF model corrected for model systematic errors
(Reforecast Project).
Image courtesy of Klaus Weickmann. See Waliser et
al. (BAMS, 2006)
27
But more work needs to be done!
28
Summary

• Most work on intraseasonal prediction has
  concentrated on the MJO, for good reason!
• Empirical MJO prediction schemes, of various
  sorts, provide skill in the region of the
  Asian-Australian monsoon out to about 20 days.
• For a benchmark statistical prediction, we
  suggest use of the Real-time Multivariate MJO
  (RMM) indices, and lagged linear regression.
• RMM phases/phase space are also very useful for
  impact studies (e.g. impact on rainfall), and for
  diagnosing the state of the MJO in dynamical
  model forecasts.
• Through our increased focus on the MJO, as
  allowed by these simple diagnostics, further
  improvements in the dynamical models will
  hopefully result (but that requires our full
  encouragement of model developers!).

29
THE END
30
(No Transcript)