Calibrating ensemble weather forecasts for warnings of extreme weather events

K. Ylinen (1) and J. Kilpinen (1)(1) Finnish Meteorological Institute, Helsinki, Finland

ILMATIETEEN LAITOSMETEOROLOGISKA INSTITUTETFINNISH METEOROLOGICAL INSTITUTE

Contact: kaisa.ylinen@fmi.fi

Finnish Meteorological Institute

P.O.Box 503, FI-00101 Helsinki, Finland

EMS Annual Meeting: European Conference for Applied Meteorology and Climatology 2017

4-8 September 2017 | Dublin, Ireland

A. INTRODUCTION B. GOALS

Finnish Meteorological Institute (FMI) is participating in

the EU-project I-REACT (www.i-react.eu) in which one

objective is to improve European level extreme weather

event detection. I-REACT aims to develop a European-

wide platform to integrate emergency management data

coming from multiple sources. The proposed system will

be targeted to public administration authorities, private

companies, as well as citizens in order to effectively

prevent and react against natural disasters.

E. CONCLUSIONS

FMI will provide the forecasted occurrence risk maps for

extreme weather events in terms of probabilities using

ECMWF (51 members, 16 km) and GLAMEPS (52 mbrs, 8 km)

ensemble models with lead times from a few hours to two

weeks for whole European region. Since ensemble forecasts

tend to be underdispersive and biased they are calibrated with

statistical methods. Calibrated ensemble forecasts will be used

also by other project partners who are providing wild fire and

heat wave hazard mapping services to the I-REACT system.

D. RESULTS

I-REACT project aims to improve prediction and management of natural disasters caused by extreme weather related events.

FMI develops and provides probabilistic weather forecasts of high-impact weather using ensemble models which are

calibrated to get more reliable forecasts.

Verification results (Fig 1 and 2) show that calibration methods used for gusts and temperature improve these ensemble

forecasts: especially spread correlates better with skill (RMSE), and also forecast skill is slightly improved after calibration.

Calibration methods

Temperature – normal ditribution

mean = a + b·MEAN + c·ELEV

stdev = exp (d + e·log(STD) + f·log(max(1,ELEV))

Wind speed and gust – box-cox t-distribution

mu (median) = a + b·MEAN + c·ELEV

sigma (variance) = exp (d + e·log(STD) + f·log(max(1,ELEV))

nu (skewness) = g + h·MEAN

tau (kurtosis) = exp(i)

Calibration coefficients

Estimated by training statistical

models using 30 most recent

point observations and

forecasts

For each lead time and for each

forecast cycle independently

Common for whole region

Updated once a week

Probabilistic

forecasts of extreme

weather events

Strong winds/gusts

Extreme high/low

temperatures

Heavy rainfall

Fig 1: Verification results for raw (blue) and calibrated (orange)

ECMWF-ENS maximum wind gust (3 hour) forecasts for lead

times from 3 to 144 hours. Verification period is June 2017

(00UTC analysis times), and verification includes all European

stations.

Fig 2: Verification results for raw (blue) and calibrated (orange)

ECMWF-ENS 2 meters temperature forecasts for lead times

from 3 to 360 hours. Verification period is June 2017 (00UTC

analysis times), and verification includes all European stations.

Fig 3: Probabilistic weather forecasts of extreme high

temperatures (left) and strong gusts (right). Forecasts are

made by using calibrated ECMWF-ENS model.

Fig 4: Fire Weather Index

(FWI) forecast which is

computed by Meteosim in

I-REACT project. FWI is

calculated using calibrated

ensemble weather

forecasts provided by FMI.

FMI

For a good ensemble system, ensemble spread and root-mean-square-error (RMSE) of ensemble mean should be positively

correlated on average.

Improving

Resilience to

Emergencies

through

Advanced

Cyber

Technologies

(MEAN~ensemble mean, ELEV~model elevation, STD~ensemble standard deviation, a-i~constants/coefficients)

C. METHODS

Fire Danger Classes FWI ranges

Very low < 5.2

Low 5.2 – 11.2

Moderate 11.2 – 21.3

High 21.3 – 38.0

Very high 38.0 – 50.0

Extreme >= 50.0

This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 700256.