Dealing with uncertainties in seasonal predictions Lauriane Batt - PowerPoint PPT Presentation

Dealing with uncertainties in seasonal predictions Lauriane Batté (CNRM, UMR 3589 Météo-France & CNRS, Université de Toulouse, France)

Introduction – Sources of uncertainty Conceptual illustration : Uncertainties in weather predictions ■ Figure 2 from Slingo and Palmer (2011) : illustration of sources of uncertainty in a probabilistic weather forecast L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 2

Introduction – Sources of uncertainty But in seasonal forecasts, there are additional sources of ■ uncertainty Figure 8 from Slingo and Palmer (2011) : illustration of sources of uncertainty in a probabilistic seasonal forecast with (a) model biases and (b) a changing climate L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 3

Introduction – Sources of uncertainty Goal of this lecture : ― Provide an overview of the different sources of uncertainty in seasonal forecasting Discuss some strategies used in state-of-the-art seasonal ― forecasting systems to deal with these uncertainties L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 4

Lecture outline Dealing with uncertainties in initial conditions ■ Dealing with uncertainties in numerical models ■ ― Multi-model approach ― Stochastic perturbations Dealing with uncertainties in seasonal forecast ■ evaluations Communicating uncertainties in seasonal forecasts ■ L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 5

Lecture outline Dealing with uncertainties in initial conditions ■ Dealing with uncertainties in numerical models Multi-model approach Stochastic perturbations Dealing with uncertainties in seasonal forecast evaluations Communicating uncertainties in seasonal forecasts L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 6

The Lorenz attractor (1963) Lorenz (1963) : Introduction of ■ chaos theory in meteorology Very simple model (non-linear ■ equations) Small errors in initial conditions ■ could lead to very large uncertainties in the time evolution on the Lorenz attractor Depending on the initial phase, the ■ growth of uncertainty (and hence predictability) differs greatly. Limits of predictability in a ■ deterministic framework : typically 10-15 days L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 7

Consequence : ensemble prediction Probabilistic weather forecasts : generated with small random ■ perturbations to the atmospheric initial conditions Conversely, when dynamical seasonal forecasts were first ■ developed, these were constructed as ensemble forecasts L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 8

Consequence : ensemble prediction Probabilistic weather forecasts : generated with small random ■ perturbations to the atmospheric initial conditions Conversely, when dynamical seasonal forecasts were first ■ developed, these were constructed as ensemble forecasts L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 9

Consequence : ensemble prediction Probabilistic weather forecasts : generated with small random ■ perturbations to the atmospheric initial conditions Conversely, when dynamical seasonal forecasts were first ■ developed, these were constructed as ensemble forecasts L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 10

Consequence : ensemble prediction Global reanalyses for the atmosphere, land, ocean provide initial ■ conditions over a range of past years ; corresponding analyses are used for real time initialization Ensemble generation techniques for initialization vary depending ■ on the institute, but generally use one of the following: Lagged initialization : ( Hoffman and Kalnay, 1983) ― ensemble members are initialized using different sets of initial conditions separated by 6 hours, one day, one week… or combinations of these for the atmosphere / ocean ― Initial condition perturbation : (Kalnay, 2003) atmosphere or ocean (re)analysis + small perturbation Ensemble assimilation : similar to the previous method, but ― members directly derived from the members of an ensemble assimilation technique L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 11

Consequence : ensemble prediction Examples : ■ ― ECMWF SEAS5: atmosphere and some land fields are perturbed using EDA perturbations from 2015, as well as leading singular vector perturbations ; ocean fields are from a 5-member OCEAN5 analysis + SST pentad perturbations (Johnson et al. 2019) CFSv2: lagged initialization with 4 runs per day every five ― days for the 9-month forecasts, 1 run per day for 1-season forecasts (Saha et al. 2014) Météo-France System 6: lagged initialization with start dates ― on the 20th, 25th of the previous month, 1 control member on the 1st L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 12

Lecture outline Dealing with uncertainties in initial conditions Dealing with uncertainties in numerical models ■ ― Multi-model approach ― Stochastic perturbations Dealing with uncertainties in seasonal forecast evaluations Communicating uncertainties in seasonal forecasts L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 13

Uncertainties in numerical models Example: CNRM-CM model co-developed by CNRM and ■ CERFACS (Voldoire et al., 2019) Atmosphere: ARPEGE Climat climate model, typically run at resolutions ~1.4° (~0.5° in System 6) Land surface: SURFEX interface Ocean: NEMO v3.6 on ORCA1 tripolar grid Coupler: OASIS MCT L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 14

Uncertainties in numerical models Numerical models are implemented on finite grids ■ → numerical approximations of the equations defining the time evolution of physical fields (e.g. Navier-Stokes equations for ocean and atmosphere) : time stepping, splitting of integration of seperate tendencies... →sub-grid scale phenomena often need to be parameterized in GCMs (e.g. triggering of convection…) → example : lower resolution models have a coarser topography and don’t represent well the impact of orography on large-scale flow L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 15

Uncertainties in numerical models Coupling different model components inevitably leads to further ■ sources of model uncertainty Representing fluxes between components ― ― Coupling frequency of GCMs is restricted by computational costs Limited availability of reference data (field campaigns) ― L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 16

Uncertainties in numerical models These model limitations inevitably lead to model-dependent and ■ flow-dependent errors that are difficult to correct a posteriori in seasonal forecasts So how can we deal with these sources of uncertainty? ■ Two strategies discussed here: ― Multi-model approach: use several models as a means of quantifying errors related to model choices Stochastic methods: introduce in-run perturbations ― accounting for model error L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 17

Multi-model approach Seminal papers: Krishnamurti et al. 1999 & 2000, Doblas-Reyes ■ et al. 2000, Hagedorn et al. 2005 Simple idea: combining ensemble forecasts from different, ■ independant models as a way of estimating the uncertainty resulting from model error 3 straightforward ways to construct a multi-model ensemble: ■ Equally weighted members (Hagedorn et al. 2005) ― Multi-model mean (equally weighted models) ― Weighted ensemble, with weights depending on model ― performance for given criteria over the hindcast period L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 18

Multi-model mean Assumption: no particular ■ model is more likely to represent the truth than any other in the multi- model Works well if levels of ■ performance are similar Fig. 3 from Mishra et al. 2019 showing at a gridpoint level the system with highest correlation, and correlation value, for EUROSIP hindcasts for DJF and JJA at lead times 2-4 months. L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 19

Weighted ensemble Several methods to determine weights have been applied in ■ past studies: ― Minimization of Ignorance score (Weigel et al. 2008) ― Bayesian approaches (e.g. forecast assimilation, Stephenson et al. 2005) Multiple linear regression techniques ― Using correlation as weights (Mishra et al. 2019) ― Due to very short verification periods, and some co-linearity ■ between the different forecasts, there is a large uncertainty in the weights derived from such techniques. To avoid over-fitting of some techniques, cross-validation is ■ necessary, and if possible, separating learning and verification periods. L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 20

Some results (Batté and Déqué 2011, ENSEMBLES project) Fig. 6 from Batté and Déqué 2011 showing the RMSE vs ensemble spread of single models and multi-model ensemble (equal weights) for the ENSEMBLES project 1960- 2005 seasonal hindcasts for JJA precipitation over West Africa (a) and DJF precipitation over southern Africa (b) L. Batté - Uncertainties - CITES 2019 School (Moscow, 31 May 2019) Page 21