Slide 1 - IRI - Columbia University

Document Sample
Slide 1 - IRI - Columbia University Powered By Docstoc
					   Seasonal Climate Forecast
               Sylwia Trzaska
International Institute for Climate and Society
    Earth Institute at Columbia University
                New York, USA

Basics of seasonal climate forecast
      - sources of uncertainty
Main methods
      - statistical
      - General Circulation Models
IRI Forecast system and products
Forecast verification
     Seasonal Climate Forecast
Global Seasonal Forecasts

                        Seasonal Climate Outlook Forums


                               PRES-AO (11)
                                              GHACOF (22)

                                   PRES-AC (3)

                                                 Consensus Regional Forecast

                                      SARCOF (11)
            Example of climate information

Historical record of temperature
in South Africa

   Climate is not a steady state
   Varies on a range of scales
                  Scales and uncertainties                                        Initial &
                                                        Initial &               Composition
              Observed     Initial & Projected         Projected
                State     State of Atmosphere        State of Ocean             Climate Change


                                                           Under research

  adapted from CLIVAR    Time Scale, Spatial Scale
  courtesy L. Goddard
Bases of Seasonal Forecast
   SST influence on Tropical atmosphere

 Modifications of boundary conditions (SST, Land
   surface) can influence weather characteristics
   (amplitude of an event, its persistence etc.,)
 Thus the characteristics of the season
      Bases of seasonal forecast – IC and BC

Example: predict the trajectory of an arrow
We do know the physical laws that govern the movement of the dart so what
determines the arrival point?

The exact conditions at the departure (the exact departure point, the exact
direction of the throw, the initial speed etc)
The conditons that can influence its trajectory (ex. Draft that could deflect the

The farther the target the more difficult the estimation of the arrival point

Weather forecast – horizon <10 days, depend mostly on the Initial Conditions (IC)
Seasonal forecats – horizon > month, depend mostly on the external influences
(Boundary Conditions – BC); SST are the key factor

      Uncertainty on Initial Conditions
      Uncertainty on the evolution of the system
  Bases of seasonal forecast – sources of
 Uncertainty     in Initial Conditions
Estimated from observed data, mostly
surface stations, distributed unevenly,
with vast areas without data

           Surface observations density

                                          Satellite data do provide some
                                          information but need calibration
                                          against observed data

                                          Solution : ensemble technique
                                          With perturbed IC
                                    Ensemble Technique
                                                                            Deterministic Forecast
Uncertainty on Initial Conditions             Uncertainty on the forecats


 Met Analyses

  Bases of Seasonal Forecast – sources of
  Uncertainty on the evolution of the system
  How do we forecast the future state of the climate?
  •Statistical Methods
  •Dynamical Methods

1. Statistical Methods: identify statistical relationships between
   variables in the past
 Ex. Relationship between a rainfall             Ex. 3 SST indices used to forecast the
 indexin E Africa and Nino3.4                    total JAS rainfall in the Sahel

Pb. The set of predictors and relationships based on the past
There are instabilities in the teleconections
  Bases of Seasonal Forecast – sources of
2. Dynamical methods: Numerical Models of Climate System

                   General Circulation Models
Attempt to quantitavely represent all the proccesses in the climate system

 • Numerical Models
 • Based on well known physical equations
 • Spatio-temporal discretisations: the quantities are
 computed for a 3D box with a given time step
             General Circulation Models

                                         Constraints on computing
                                         = constraintes on the resolytion

                                         Typical GCM resolution for climate
                                         purposes 250x250km
                                         Time step 15min

Sources of Errors:
 Scale of numerous phenomena < model grid scale
(ex. clouds)
 Models of different subsytems (ex. atmosphere,
ocean) developped separetly, assuming the other
component perfectly known/represented
                A part of the world according to a GCM...

                                                            red: land
                                                           white: water

                                                           1 box=500 km

                    Can you recognize the region?
Courtesy Prof Deliang Chen, Goteborgs University, Sweden

                                                1 box=250 km

                                                           1 box=180 km

                                  1 box=110 km

Courtesy Prof Deliang Chen, Goteborgs University, Sweden
   Changing Scales - Downscaling

 dynamical                                                  Statistical
downscaling                                                downscaling

Courtesy Prof Deliang Chen, Goteborgs University, Sweden
    Dynamical Systems for Seasonal Forecast
•Systems where the moatmospgeric and ocean modelas are coupled i.e.
Communicate with each other in both directions ( the state of the atmosphere is
modified based on ocean conditions and the state of the ocean is influenced by the
atmosphere )

Ex. 1 Tier System, ex. ECMWF
          the system (atm and oc) is initialised the is left to evolve on its own

•Uncoupled systems where the evolution of one of the components is prescribed
(ex. evolution of the SST) and influences the other component (atmosphere)without
being influenced back

Ex. 2 Tier System, ex. IRI
          The evolution of the SST over the next 12 months is forecasted externally
          - Tropical Pacific SST forecasted using Cane-Zebiak model
          - SST in other basins :      1/ persistance
                                       2/ statistical methods

Finally : the forecatst use multimodel-ensembling and each model is
integrated with several perturbed initial conditions
    IRI operational Seasonal Climate
 IRI is issuing operationally each month (around 15th of each month)
the seasonal forecast for a number of forthcoming seasons

     M       J      J       A      S      O       N

  Available on the web as maps for
 RR and temp
  some data available for partners
IRI Dynamical Climate Forecast System
                           2-tier system
     OCEAN                 ATMOSPHERE
       PERSISTED              GLOBAL
        GLOBAL              ATMOSPHERIC
                                             10 Persisted
        ANOMALY                              24    SST
                             ECPC(Scripps)   24
                                                3 Mo. lead
                                             10                   POST
  FORECAST SST              ECHAM4.5(MPI)
   TROP. PACIFIC             CCM3.x(NCAR)
(multi-models, dynamical                                       MULTIMODEL
    and statistical)           NCEP(MRF9)                      ENSEMBLING
                                             24 Forecast
 TROP. ATL, INDIAN            NSIPP(NASA)    24     SST
     (statistical)                           30
                                   COLA2     12 3/6 Mo. lead
 (damped persistence)               GFDL
  Post-processing of multi-ensemble multi-model

•Compute the seasonal mean (ex. JAS) for different variables (precipitation,
temperature) in each grid point
•Compute the main statistics (mean, median, std)
•Comparison with climatological distribution => estimation of probabilites
Past climatology of the ensemble


     33%    33%     33%                                  High probability of rainfall in the above normal category

                      Automatic Corrections
                      • Each model grid point is weight in function of the past
                      • Spatial smoothing applied to the ewights
                      • More recently – baysian approach
ENSO Forecasts
IRI Seasonal Forecasts
New forecast format - probability of
                          ECHAM4.5 2m Temperature:
                          JFM 1983 – El Nino

                        Climatology based on 1950-2001 period.

                                     (Courtesy Lisa Goddard)
         Forecast Verification

 It is important to provide the user with some
assessment of the quality of the forecast

   Quality – degree to which a forecast corresponds to observations

   Value – degree to which a forecast helps the decision maker
           realize incremental economic (other) benefit

          • High quality forecast – little value: clear sky over Sahara
          • Low quality forecast - high value : prediction of fog at
          airports => even if overcast (false alarms)
      Assessing Forecast Quality
  Probabilistic forecasts address the two fundamental
             • What is going to happen?
             • How confident can we be that it is
             going to happen?
  Both these aspects require verification
  There is no single measure that gives a comprehensive
  summary of forecast quality.

NB. As soon as a forecast is expressed probabilistically, all possible outcomes
are forecast
Probabilistic forecasts cannot be right or wrong, but they can be over- or
    Assessing Forecast Quality
•How do we decide whether a forecast was “correct”?
•How do we decide whether a set of forecasts is correct
 consistently enough to be considered “good”?
•What is the difference between an “accurate” forecast, and
 a “skilful” forecast, and a “reliable” forecast?

  Accurate – close to observations
  Reliable – average agreement with observations
  Skilful – average accuracy with respect to reference
   forecast (climatology, persistence)

 Skill score – comparison with a reference forecast
 Verification score – comparison with observations
   Probabilistic Forecast Verification

Whenever a forecaster says there is a high probability of rain
tomorrow, it should rain more frequently than when the forecaster
says there is a low probability of rain.

   A consistency between the forecast probabilities for an event
   and the observed relative frequencies of this event.

   If the proportion of times that the event occurs is the same as
   the forecast probability for all values of the prior probability, the
   system is reliable (or well calibrated).

   Murphy, A. H., 1973: A new vector partition of the probability score. Journal of Applied
   Meteorology, 12, 595–600.
   Murphy, A. H., 1993: What is a good forecast? An essay on the nature of goodness in
   weather forecasting. Weather and Forecasting, 8,         281–293.
Reliability Diagram
Multimodel ensemble verification
 Major Goal of
Forecasts should
“mean what they say”.

                        Courtesy Tony Barnston
    Probabilistic Forecast verification
                 method 2
       Relative Operating Characteristics (ROC)

•   Mainly addresses the following questions:
    – For how many of the events were warnings correctly
    – For how many of the non-events were warnings
       incorrectly provided?
•   ROC measures how good forecasts are in the
    context of a very simple decision-making model
    from the perspective off the user.
•   ROC recognizes that forecast quality cannot be
    measured by a single number
ENSO Forecasts

Shared By: