Support for the 2014 Olympic Games in Sotchi Pierre Eckert MeteoSwiss, Geneva

Federal Department of Home Affairs FDHAFederal Office of Meteorology and Climatology MeteoSwiss

Support for the 2014 Olympic Games in Sotchi

Pierre EckertMeteoSwiss, Geneva

COSMO WG4 coordinator« Interpretation and applications »

COSMO General meeting, September 2010

2 Sotchi Olympic Games, General Introduction ¦ COSMO General Meeting September 2010Pierre.Eckert[at]meteoswiss.ch

Plan of the session

• General introduction (P. Eckert)• Postprocessing / statistical downscaling• Some methods used in Vancouver 2010

• Experiences from Torino 2006 (M. Milelli)• Input from Roshydromet (I. Rosinkina, G. Rivin,…)

• Know-how in postprocessing• Planed organisation / setup of measurements

• Elements of discussion (for further treatment)• Setup of 0.5-2 km model, incl. assimilation• Implementation of a probabilistic model (EPS)• Connection with demonstration project• Role of the COSMO w.r. to other collaborations

• Definition of WG4 working packages

3

Enhanced observational network; Nowcasting tools; Regional data assimilation; High-resolution NWP models and EPS; Meso-scale verification system; Means of NWP output interpretation and

delivery (new parameters and products, visualization etc); postprocessing;

Training

Primary meteorological needs for Sochi-2014:


Postprocessing

• Derived fields: pressure levels, PV, radar reflectivity,…• Generation of products: TV, Internet,…• Diagnostics: turbulence, icing, snowfall limit,…

• Local adaptation, downscaling• Statistical downscaling (correction of model with

observations)• Blending (mixture of model output and observations

(gridded), INCA,…)• Downstream models (1d, 2d, 3d,…)


Perfect Prog, MOS

Use two sets of historical data:1. The predictand = the local element you want to predict:

temperature at Sotchi, occurrence of fog on the downhill slope,…

2. The predictors = a bunch of model parameters: pressure, instability indices, 850 hPa temperature, winds,…It is allowed to take recent observations of the predicand as predictor.

Correlations (regression, discriminance,…) between the predictand and the predictors are computed. Often the predictors are selected by significance.

Kalman filtering is probably a subclass


Non linear methods

• The same data sets can be treated with non linear methods• Neural networks• Boosting• …

• Instead of defining hyperplanes in the predictor space, arbitrary shapes can be found.

• The selection of predictors, the choice of an optimal separation surface and the computation of coefficients is called “learning”


Classification

• A set of fixed meteorological situations is defined.• Every country has several such classifications• They are usually correlated to sensible weather (in

the situation 7b, the sun is shining in 90% of the corresponding days,…), in situation SWa there is an 80% chance to get hill fog over the downhill slope,…


Classification and interpretation

Luganorain > 1 mm/24h

Luganorain > 10 mm/24h %


Analogs

• This method looks for the n situations in the past which are closest to a given forecast according to some distance.

• A statistics on the weather elements corresponding to these n situations is then made.

• As with the classifier, it is possible that the closest situation is far away from the presented situation.


Statistical Adaptation for COSMO

COSMO General Meeting 2010

Vanessa Stauch


calibration with Kalman Filter

>> recursive estimation of forecast error (prediction – correction)

>> requires online observations

>> can be used quasi-instantaneously (no large historical database)

>> cannot predict fast changes (assumption of persistent error for each fcst)

>> suitable for a subset of parameters (normally distributed errors)


Kalman Filter @ MeteoSwiss

operational:

T2m, TD2m

for

COSMO-LEPS mean

COSMO-7

COSMO-2

IFS

in preparation:

FF10m, TW2m, RH2m

for

COSMO-LEPS mean

COSMO-7

COSMO-2

IFS


T2m predictions

COSMO-7

COSMO-2COSMO-2

KFC7

COSMO-7

KFC2

COSMO-7COSMO-7

performance?


benefit COSMO-2 vs COSMO-7?

04.04.08 – 31.10.08 RMSE RMSE

All ANETZ stations 11.0 % 4.0 %Low ANETZ stations 8.8 % 5.2 %High ANETZ stations 13.0 % 3.2 %

04.04.08 – 31.10.08 STD STD

All ANETZ stations 12.3 % 4.0 %Low ANETZ stations 12.7 % 5.2 %

High ANETZ stations 12.6 % 3.2 %

C2 vs C7 C2-KF vs C7-KF

=

Differences between COSMO-2 and COSMO-7 with KF smaller but still significant. Bias in KF predictions totally removed


Short term Kalman filter for radiation

Short-term correction: based on previous hour and every new obs

» exploits temporal autocorrelation of the error with the Kalman filter» corrects a few hours only» also beneficial for temperature forecasts

Zurich, 01.07.-10.07.2007solar heat gain for south orientation (derived from global radiation)


Calibration with multiple regression MOS

>> estimation of multiple linear regression models

>> requires large historical database (observations and forecasts)

>> models can be „arbitrarily“ complicated (provided the data)

>> possibly less adaptive than the KF (constant regression parameters)

>> suitable for a larger subset of parameters (compared to KF)


COSMO-7 vs COSMO-2

COSMO-7 COSMO-2 (03)Chasseral (CHA) 56 > 44Evionnaz (EVI) 90 < 99Gütsch (GUE) 55 > 45Oron (ORO) 53 < 59Piz Martegnas (PMA) 51 > 45Schaffhausen (SHA) 54 > 52Uetliberg (UEB) 76 > 69

rRMSE (%) für 1-24h, period 01.09.08 – 31.03.09

CHA

EVI

GUEPMA

SHA

OROUEB


effect on MOS-postprocessing

COSMO-7 MOS

COSMO-2 (03) MOS

Chasseral (CHA) 34 > 32Evionnaz (EVI) 78 > 77Gütsch (GUE) 45 > 39Oron (ORO) 59 > 47Piz Martegnas (PMA) 69 > 42Schaffhausen (SHA) 85 > 77Uetliberg (UEB) 59 > 54

rRMSE (%) für 1-24h, period 01.09.08 – 31.03.09

CHA

EVI

GUEPMA

SHA

OROUEB


plans for COSMO-MOS

>> development of a regression-based model output statistics system

>> target parameter:

wind speed and direction, sunshine duration, global radiation

>> using information of COSMO-7, COSMO-2 and COSMO-LEPS

>> project duration 08.2010 – 08.2012


Summary

Statistical postprocessing profits from a better NWP input model

„dynamical downscaling“ does not replace statistical adaptation to

local observations (in particular if results being verified against

those)

Long time series of model forecasts and observations (≥ 2 years)

are prerequisite for the development of a robust statistical model


1d, 2d, 3d models

• It is also possible to feed 1d, 2d, 3d models forced by the 3d (4d) model.

• Ex. Fog model: soil model, a lot of levels in the few 10’s of meters of the atmosphere, aerosols,…

• Should ideally be incorporated into the full model, but can be expensive.


Local 2d and 3d models for the 2010 Vancouver Olympic Games

COSMO General Meeting 2010

Thanks to André Méthod, CMC

23

Real-Time Experimental Land Surface System Real-Time Experimental Land Surface System for the 2010 Vancouver Games for the 2010 Vancouver Games

with contributions from:

Maria Abrahamowicz, Bernard Bilodeau, Marco Carrera, Nathalie Gauthier, Lily Ioannidou, Alain Patoine, et

Sylvie Leroyer

Natacha Bernier Linying Tong

andStéphane Bélair

SLIDE 1

24

Concept of external land surface modeling (again!)

ATMOSMODEL

3D INTEGRATION

ExternalLand SurfaceModel

With horizontal resolution as high as that of surface databases (e.g., 100 m)

ATMOSPHERIC FORCING at FIRST ATMOS. MODEL LEVEL (T, q, U, V)

2D INTEGRATION

Computational cost of off-line surface modeling system is much less than an integration of the atmospheric model

ATMOSPHERIC FORCING at SURFACE (RADIATION andPRECIPITATION)

LOW-RES

HIGH-RES

SLIDE 2

25

Applications to the 2010 Vancouver Games:Two surface systems: “2D” and “Point”1400 x 1800 computational grid (100-m grid size)

USA

VAN

Whistler

Blackcomb

Callaghan

VANCOUVERCypress

Bowl

SLIDE 3

26

Experimental real-time “2D” land surface system

ANALYSIS / ASSIMILATION

FORECAST

24-h open-loop run

Geophysicalfields

00 UTCInitial timeSurface analysis

REG-15 12Z (12-18h)

REG-15 00Z (6-18h)

REG-15 12Z (6-12h)

ATMOSPHERIC FORCING

24-h open-loop runfor next day

96-h forecast run

REG-15 00Z (0-48h)Av. at 03Z

48h

GLB-33 00Z (48-96h)Av. at 06Z

SLIDE 4

27

Experimental real-time “point” land surface system

ANALYSIS / ASSIMILATION

24-h background run

Geophysicalfields

00 UTCInitial timeSurface analysis

Snow obs

ATMOSPHERIC FORCING

24-h background runfor next day

96-h forecast run

SCREEN-LEVEL OBS +MODEL FORCING

FORECAST

SLIDE 5

REG-15 00Z (0-48h)Av. at 03Z

48h

GLB-33 00Z (48-96h)Av. at 06Z

28

Two-dimensional snow analysis against surface observations

(Bernier et al. 2010, part I)

Close relationship with height of observations and of model outputs, ... but not always...

SLIDE 6

29

Verification of “point” snow analysis at VOC

VOC Blackcomb Mt. Base

2008

REG-OP (15 km)

“POINT”

OBS

2D-100m

LAM-OP (2.5 km)

Atmospheric forcing (e.g., precipitation phase) is of crucial importance for the 2D system (without assimilation of surface snow obs)

As could be expected, “point” system is right on target (because of the asssimilation of surface snow data)

SLIDE 7

33

List of products

• Last 10 days meteograms (forcing + screen-level diagnostics from surface system)

• Last 10 days surfacegrams (surface prognostic variables – focus on snow conditions)

• Next 4 days meteograms (forcing + screen-level diagnostics from surface system)

• Next 4 days surfacegrams (surface prognostic variables – focus on snow conditions)

SLIDE 11

34

Examples of Meteograms and “Surfacegrams”

SLIDE 12

1.0 km

WhistlerWhistler

VancouverVancouver

15 km

2.5 km

High resolution Numerical Weather Prediction Systems High resolution Numerical Weather Prediction Systems for the Vancouver 2010 Winter Olympics and for the Vancouver 2010 Winter Olympics and Paralympics GamesParalympics Games

A. Erfani, B. Denis, A. GigA. Erfani, B. Denis, A. Giguèreuère, , N. McLennan, N. McLennan, A. Plante, L. Tong, A. Plante, L. Tong,

Environment Canada / MSC/ DevelopmentEnvironment Canada / MSC/ Development

S. BS. Bélair, élair, M. Charron,M. Charron, J. Mailhot, J. Mailhot, R. McTaggart-Cowan, J. MilbrandtR. McTaggart-Cowan, J. Milbrandt

Environment Canada / Meteorological Research Environment Canada / Meteorological Research DivisionDivision

High Resolution Prediction System - cascading

Growing Orography• 3 h → LAM_2.5• 1 h →LAM_1.0

00 06 18 00 06 12 18 0000 06 12 18 12UTC

16 22 10 16 22 04 10 1616 22 04 10 04PST

00 06 18 00 06 12 18 0000 06 12 18 12UTC

16 22 10 16 22 04 10 1616 22 04 10 04PSTREGIONAL R100 (48hrs)

LAM_15 (39hrs)

LAM_2.5 (33hrs)

LAM_1.0 (19hrs)

06Z.

11Z. 06Z

15Z

REGIONAL R100 (48hrs)

LAM_15 (39hrs)

LAM_2.5 (33hrs)

LAM_1.0 (19hrs)

06Z.

11Z. 06Z

15Z

00Z.

00 06 18 00 06 12 18 0000 06 12 18 12UTC

16 22 10 16 22 04 10 1616 22 04 10 04PST

00 06 18 00 06 12 18 0000 06 12 18 12UTC

16 22 10 16 22 04 10 1616 22 04 10 04PSTREGIONAL R112 (48hrs)

LAM_15 (36hrs)

LAM_2.5 (33hrs)

LAM_1.0 (19hrs)

15Z.

20Z. 15Z

00Z

12Z.

LAM 2.5km Product period

LAM 1.0km Product period

Available to forecasters:

• by 7:00 a.m. local (for the morning briefing)

• by 12:00 noon local (for afternoon briefing)

• Based on Olympic forecasters’ feedback:- products, display format,…

• Easy display (Weather Viewer)• Comprehensive list of model outputs:

- 2D maps, time series at stations, vertical soundings and cross-sections

• Products available for evaluation by support desk and briefings

Customized output packageCustomized output package

2D maps:• Screen-level potential temperature• Screen-level relative humidity• 10-m winds• Wind gusts (gust estimates, minimum, maximum)• Standard deviations of 10-m wind speed and direction • Accumulated precipitation types (liquid / freezing / snow /

frozen)• Precipitation accumulation (liquid / solid / total)• Precipitation rate (liquid / solid / total)• Snow/liquid ratio• Cloud cover (high/ mid/ low)• Cloud base height• Visibility (through fog, rain, snow)• Freezing level (0C isotherm)• Snow level• Wind chill factor


TemperatureTemperatureDewpoint, RHDewpoint, RH

Surface windsSurface winds

Cloud cover:Cloud cover:Low, mid, highLow, mid, highLow level windsLow level winds Wind gustsWind gusts

WindchillWindchillCloud baseCloud base

QPF: QPF: /1h, /3h,/1h, /3h, /6h, /6h, cumulativecumulative

QPF QPF by type:by type:cumul. , instant.cumul. , instant.

2-D maps2-D maps(1 km)(1 km)

Visibility: Visibility: fog, rain, snow, fog, rain, snow, resultantresultant


MeteogramsMeteograms(1 km)(1 km)

General WxGeneral WxT, TT, Tdd etc. etc.

Wind and GustsWind and Gusts

PrecipitationPrecipitationPCP RatesPCP Rates

Clouds and visibilityClouds and visibility

SnowSnow


General weather:low-level temperature,

cloud cover,

total precipitation,

wind speed and direction

1-km LAM model

Callaghan Valley (VOD)


Precipitation Precipitation AmountsAmountsRatesRates

High Resolution Prediction System - Multi-model Meteograms

Thank You!Thank You!


Plan of the session

• General introduction (P. Eckert)• Postprocessing / statistical downscaling• Some methods used in Vancouver 2010

• Experiences from Torino 2006 (M. Milelli)• Input from Roshydromet (I. Rosinkina, G. Rivin,…)

• Know-how in postprocessing• Planed organisation / setup of measurements

• Elements of discussion (for further treatment)• Setup of 0.5-2 km model, incl. assimilation• Implementation of a probabilistic model (EPS)• Connection with demonstration project• Role of the COSMO w.r. to other collaborations

• Definition of WG4 working packages

Documents

Support for the 2014 Olympic Games in Sotchi Pierre Eckert MeteoSwiss, Geneva