ECMWF training course, 2006: Predictability on the seasonal timescale Predictability on the Seasonal Timescale Tim Stockdale and Franco Molteni Seasonal

ECMWF training course, 2006: Predictability on the seasonal timescale

Predictability on the Seasonal Timescale

Tim Stockdale and Franco Molteni

Seasonal Forecast Section

European Centre for Medium-Range Weather Forecasts


Contents

Sources and limits of seasonal predictability El Nino is predictable, and has a major impact on the global atmosphere Many other factors, not all understood; limits of empirical methods

Seasonal forecasting with ocean-atmosphere GCMs Benefits of an ensemble forecast with a comprehensive numerical model Outline of ECMWF system and handling of model bias The problem of model error

Performance of the ECMWF system El Nino forecasts Atmosphere behaviour Validation and operational use


Basic ideas of seasonal predictability

Internal variability of atmosphere is mostly unpredictable

Boundary conditions and “external” forcings may give partial/probabilistic predictability

The boundary/external forcing must be predictable!Either slowly changing (eg CO2 trend, maybe decadal

variability)Or dynamically predictable (eg El Nino)Or have at least a moderate persistence timescale (eg large

soil moisture anomalies)


Sources of seasonal predictability

KNOWN TO BE IMPORTANT: El Nino variability - biggest single signal Other tropical ocean SST - important regional impacts Local land surface conditions - e.g. soil moisture in spring Climate change (all forms) - especially important in mid-latitudes

OTHER FACTORS: Mid-latitude ocean temperatures- still controversial Remote soil moisture/snow cover - not well established Volcanic eruptions - definitely important for large events Sea ice anomalies - local effects are clear Stratospheric QBO - possible tropospheric impact Dynamic memory of atmosphere - most likely on monthly

timescale Solar cycle - questionable statistical connections


Methods of seasonal forecasting

Empirical forecasting schemes Use past observational record and statistical methods Work with observed data instead of error-prone numerical models Limited number of past cases means that they work best when observed

variability is dominated by a single source of predictability A non-stationary climate is problematic

Two-tier forecast systems First predict SST anomalies (ENSO or global; dynamical or statistical) Use ensemble of atmospheric GCMs to predict global response Problems if SST variability is mainly forced by atmospheric variability

Single-tier GCM forecasts Include comprehensive range of sources of predictability Predict joint evolution of SST and atmosphere flow Estimate uncertainty of future SST, important for prob. forecasts Model errors are an issue!


ICTP hindcast of Indian summer rainfall (land points 70-95 E, 10-30N) vs. CRU data

Prescribed global observed SST mixed layer in IO, obs. SST elsewhere corr = 0.28 corr = 0.44


ICTP hindcast of Indian summer rainfall (land points 70-95 E, 10-30N) vs. CRU data

Dynamical ocean model (MICOM 2.9)in the IO, observed SST elsewhere

corr = 0.62


ECMWF coupled model (System 2)

IFS (atmosphere) TL95L40 Cy23r4, 1.875 deg grid for physics (operational in 2001) Initialized from ECMWF operational system Prognostic clouds, fully interactive soil moisture, convection, radiation, ….

HOPE (ocean) Global ocean model, but sea-ice prescribed from climatology 1 x 1 deg at mid-latitudes, 0.3 deg meridional near equator; 29 levels Initialized by improved Optimal Interpolation analysis scheme

OASIS (coupler) Coupling once per 24 hours (so no diurnal cycle in ocean) No flux correction or other constraints (except specified sea-ice)


ECMWF coupled model (System 3, soon)

IFS (atmosphere) TL159L62 Cy30r1/2, 1.125 deg grid for physics Full set of singular vectors from EPS to perturb atmosphere initial

conditions. Ocean currents coupled to atmosphere boundary layer calculations

HOPE (ocean) Essentially the same ocean model A lot of extra work in improving the ocean analyses

OASIS (coupler) Better treatment of sea-ice, but still no proper model


ECMWF forecast strategy

Initialize coupled system (see Magdalena’s talk) Aim is to start system close to reality. Accurate SST is particularly

important, plus ocean sub-surface.

Run an ensemble forecast (see Magdalena’s talk) Generate an ensemble on the 1st of each month, with perturbations to

represent the uncertainty in the initial conditions; run forecasts for 6 months

Remove systematic error Forecasts have considerable systematic error Estimate this error from a set of previous forecasts, which define the

model climatology. Model climatology is a function of date and forecast lead-time. Linear assumption is not correct, but seems to work reasonably well.


Creating the ensemble

Wind perturbations Perfect wind would give a good ocean analysis, but uncertainties are

significant. We represent these by adding perturbations to the wind used in the ocean analysis system.

SST perturbations SST uncertainty is not negligible SST perturbations added to each ensemble member at start of forecast.

Atmospheric unpredictability Atmospheric ‘noise’ soon becomes the dominant source of spread in an

ensemble forecast. This sets a fundamental limit to forecast quality. To account for uncertainties in physical parametrizations, we activate

‘stochastic physics’.


Remove systematic error

Model drift is typically comparable to signal Both SST and atmosphere fields

Forecasts are made relative to model climatology Model climate estimated from 15 years of forecasts (1987-2001), all of

which use a 5 member ensemble. Model climate has both a mean and a distribution, allowing us to estimate

eg tercile boundaries. Model climate is a function of start date and forecast lead time. EXCEPTION: Nino SST indices are bias corrected to absolute values, and

anomalies are displayed wrt a 1971-2000 climate.Implicit assumption of linearity

We implicitly assume that an anomaly in the model forecast relative to the model climate corresponds to the expected anomaly in an unbiased forecast relative to the true climate, despite differences between model and true climate.

Most of the time, this assumption seems to work pretty well.


SST bias is a function of lead time and season






Despite SST bias and other errors, anomalies in the coupled system can be remarkably similar to those obtained using observed (unbiased) SSTs …..


… and can also verify well against observations


Model errors are serious

Models have errors other than bias Eg weak variability in system 2 and recent test versions

Errors in model climate interact with errors in model variability

Specific example in Nino 4 region Impact of artificial change in mean wind stress

Our forecast errors are larger than they should be With respect to internal variability estimates and (sometimes) other

prediction systemsBenefits of multi-model ensembles (see Antje’s and Paco’s

talks) Gets round some of the model-specific error Fundamentally, though, we still need better models


Variability of zonal wind in System 2 is strongly damped, even in the early months when SST is close to observed


SST variability in Nino 3 is correspondingly weak


Experiments with an artificial time-invariant zonal wind-stress addition along the equator, designed to make the ocean mean state in the coupled system more realistic. Compared to ‘heat-flux’ experiments, the emphasis is on altering the sub-surface structure and hence sensitivity of the SST to sub-surface anomalies.

Sensitivity of forecasts to the mean state


The result is increased amplitude of SST anomalies, especially in the east Pacific, and in this case a small increase in forecast skill.

Increasing the amplitude of SST anomalies by this method is very robust – stronger upwelling in the mean state = stronger SST anomalies.

Improving skill is more sensitive to the details of the adjustment, however.


In Nino 4, System 2 has much bigger errors than its estimate of the predictability limit (red solid cf red dashed). In fact, a simple statistical model can give much better results (blue)

(The statistical model is a cross-validated linear regression on the analysed zonal mean heat content of the equatorial ocean and the intial value of Nino 4 SST. The coupled model has significantly more information available, but model errors are degrading its performance.)

Size of forecast errors: example of Nino-4


Fortunately for us, the GCM is still significantly ahead of the statistical model in Nino 3, at least for shorter lead times.


Operational seasonal forecasts

ECMWF runs a 40 member ensemble forecast Initial conditions are valid for 0Z on the 1st of a month Forecast is created typically on the 11th/12th

Forecast release date is 12Z on the 15th.

Range of operational products Plots of basic fields on web Raw data in MARS Formal dissemination of real time forecasts now available

Multi-model system now running UKMO and Meteo-France both run systems at ECMWF Full set of output is archived, derived products and plots are produced Operational multi-model products still to be decided


Example forecast products

A few examples only – see web pages for full details and assessment of skill

Note: Significance values on plots A lot of variability in seasonal mean values is due to chaos Ensembles are large enough to test whether any apparent signals are

real shifts in model the pdf We use the w-test, non-parametric, based on the rank distribution NOT related to past levels of skill

Note: Proper usage Interpretation must be done carefully Serious applications of the forecasts should use the actual values of the

forecast and climate distributions, not just visual charts. Model errors are significant, and must be accounted for in real

applications; techniques are not yet well developed.


The forecast showed in the 2004 training course ….





Tercile probabilities 15th percentile probabilities



And how do we do?

ROC score is one measure of the valid information contained within probability forecasts


Blue is good, yellow is bad

But based on 15 years, sampling error is substantial for local values, so spatial distribution is unreliable except where skill is high

Upper tercile


Precipitation is generally noisier …..






Documents

ECMWF training course, 2006: Predictability on the seasonal timescale Predictability on the Seasonal Timescale Tim Stockdale and Franco Molteni Seasonal