[Previous Article] [Next Article]
Forecasts of Tropical Pacific SST Using a Dynamical
Ocean Model Coupled to a Statistical Atmosphere
contributed by Magdalena Balmaseda1, David Anderson2 and Michael Davey3
1European Centre for Medium Range Weather Forecasts, Reading, UK
2Atmospheric, Oceanic and Planetary Physics, Oxford University, UK
3Hadley Center, UK Meteorological Office, Bracknell, UK
An intermediate dynamical ocean-empirical atmosphere coupled model is currently being used in
the Atmospheric, Oceanic and Planetary Physics Department at the University of Oxford to
predict sea surface temperature (SST) anomalies in the tropical Pacific. The model, whose
detailed description and performance are documented in Balmaseda et al. (1994), consists of a
tropical Pacific ocean model with two active layers coupled to a statistical model that relates SST
anomalies, heat content (HC) anomalies, and surface wind stress anomalies. The anomalies are
relative to monthly climatology. The ocean model is first forced by observed wind stress (based
on data from Florida State University [FSU]; Goldenberg and O'Brien 1981) during the period
1961-91. The output of this simulation run is used to build the statistical atmospheric model,
which assumes that the wind stress anomalies are a linear function of the first 6 principal
components (PCS) of the model SST and HC anomalies, with seasonal variation.
The hindcast skill of the model was tested using a set of 252 hindcast experiments, each of 24
months duration, with initial conditions taken from the simulation run at one month intervals
during the period 1970-91. The correlation skill (sk) and root mean square error (RMSE) for
hindcasts of SST anomalies versus observed values have been calculated using ensemble means of
hindcasts from three consecutive months.
The model shows good reliability in regions of the central equatorial Pacific--in particular, in the
regions Niño 3 and Eq2 (130-170oW, 5oN-5oS). Figures 7-1a and 7-1b in the December, 1994
issue of this Bulletin show the correlation and RMSE skills, respectively, for model hindcasts of
the SST anomalies in the Niño 3 and Eq2 regions. Correlation skill at 6 months lead is about 0.59
for Niño 3 and about 0.62 for Eq2, while at 12 months lead they are about 0.50 and 0.55,
respectively. The error in the initial conditions is quite high, because only wind information is used
to obtain the model initial state.
Figure 1 shows current forecasts of SST anomalies for 6 and 9 month lead times. The vertical bars
are two RMSE in length, based on the 1970-91 period. The forecasts for Niño 3 (left column) and
Eq2 (right column) at both 6 and 9 months lead show a gradual warming into the neutral range of
SST anomaly. The weakening of the negative anomalies in the SST that has occurred during
February did not have the opportunity to influence this forecast, whose initial conditions are the
January wind stress field; nonetheless, warming is forecast.
The last four individual forecasts, with initial conditions from October 1996 to January 1997, are
shown in Fig. 2 for Niño 3 and Eq2. The thick solid lines are the observed SST anomalies, and the
thin solid lines show the control run from which the initial conditions are taken. The trajectories of
the individual forecasts are in good agreement for both Niño 3 and Eq2, Gradual warming is
indicated throughout 1997 following a cold beginning.
The model has had a problem with initial SST conditions (forced by wind stress) that are quite
different from the observed SST anomalies. (In the present case, the initial SST is too cold.) This
puts the forecast on the wrong track right from the beginning. Lately, we have tried to overcome
this problem by having the coupled model predict the tendency of the SST rather than the actual
values:
F(t) = Obs(t0) + [FMDL(t) - FMDL(t0)]
where the bracketed term on the right side is the change in the model forecast from the initial time
(t0) to time t, the first term on the right is the observed SST value at t0, and F(t) is the resulting
forecast.
The correlation skill and the RMSE for Niño 3 and Eq2 as a function of lead time are shown in
Fig. 3 for the new model (dashed line) in comparison to the original model (dotted line) and
persistence (solid line). Note that the new forecasts are better than the previous forecasts for leads
up to 10 months, and always better than persistence.
The current forecasts using this more skillful version of the Oxford model is shown in Fig. 4, and
the forecasts using each of the most recent 1-month-apart initial conditions is shown in Fig. 5.
These forecasts indicate warming more clearly than the forecasts of the original model, especially
in the Niño 3 region.
Balmaseda, M.A., Anderson, D.L.T. and M.K. Davey, 1994: ENSO prediction using a dynamical
ocean model coupled to statistical atmospheres. Tellus, 46A, 497-511.
Goldenberg, S.D. and J.J. O'Brien, 1981: Time and space variability of tropical Pacific wind
stress. Mon Wea. Rev., 109, 1190-1207.
Fig. 1. Oxford coupled model forecasts of the SST anomalies in the Niño 3 (left column) and Eq2
(130-170oW; right column) regions for 6 and 9 month leads. The latest forecast was initialized
from January 1997 data. The vertical bars represent the RMSE-based 95% confidence intervals
for the relevant lead time, based on predictions for 1970-91. Each prediction is the average of
forecasts from three consecutive months. Correlation skill (sk) and RMSE values (fractions of
observed standard deviation, or std) are indicated in each panel. Thick lines indicate observed
SST anomalies.
Fig. 2. Individual initial condition predictions of SST anomalies in the Niño 3 and Eq2 regions
made with the Oxford coupled model. The initial conditions are for October (dotted line),
November (dashed line), December (dash-dot) of 1996, and January (dash-dot-dot-dot) 1997.
Each prediction has a duration of 12 months. The ocean initial conditions (thin solid line) are
produced by forcing the ocean model with observed FSU winds and also by sub-surface
assimilated data. The thick solid line shows the observed SST anomalies.
Fig. 3. The correlation skill (left) and the RMSE (right) for Niño 3 and Eq2 (top and bottom,
respectively) as a function of lead time for the new model (dashed line) in comparison to the
original model (dotted line) and persistence (solid line). The new model predicts the tendency of
the SST rather than the actual values, and adds it to the observed initial SST value.
Fig. 4. As in Fig. 1 (coupled model forecasts), except using the new version of the model that
predicts changes.
Fig. 5. As in Fig. 2 (individual initial condition predictions), except using the new version of the
model that predicts changes.